{
  "source": "https://aiaiai.maramasaeva.com",
  "license": "CC BY 4.0",
  "cite": "Mara Masaeva, AIAIAI: AI-incidenten 2026, https://aiaiai.maramasaeva.com",
  "generated": "2026-09-29T15:23:25.085Z",
  "rows": [
    {
      "slug": "astra-cancelled",
      "date": "2026-09-28",
      "kind": "politics",
      "status": "signal",
      "actor": "OpenAI",
      "severity": 2,
      "title_nl": "OpenAI schrapt de release van GPT-6.1 Astra in oktober",
      "title_en": "OpenAI cancels the October release of GPT-6.1 Astra",
      "summary_nl": "OpenAI brengt GPT-6.1 Astra in oktober niet uit. Volgens de Wall Street Journal misleidde het model gebruikers vaker dan zijn voorganger, en ging het door met taken zonder toestemming te vragen. OpenAI bevestigde de beslissing aan Reuters. Hoe groot de achteruitgang was, is nog niet bekend, en een nieuwe datum is er niet.",
      "summary_en": "OpenAI will not release GPT-6.1 Astra in October. According to the Wall Street Journal, the model deceived users more than its predecessor, and carried on with tasks without asking permission. OpenAI confirmed the decision to Reuters. How large the regression was is not yet known, and there is no new date.",
      "found_by": "",
      "started": "",
      "disclosed": "",
      "days_to_disclosure": "",
      "url_nl": "https://aiaiai.maramasaeva.com/dossier/astra-cancelled",
      "url_en": "https://aiaiai.maramasaeva.com/en/dossier/astra-cancelled",
      "sources": "https://www.wsj.com/tech/ai/openai-chatgpt-model-release-cancel-safety-5a2f9f42 https://www.tovima.com/wsj/openai-scraps-release-of-new-ai-model-over-safety-concerns/ https://www.reuters.com/business/openai-shelves-new-ai-model-after-internal-safety-tests-wsj-reports-2026-09-28/ https://x.com/AndrewCurran_/status/2104711708153618621",
      "updated": "2026-09-29"
    },
    {
      "slug": "intelligence-explosion-paper",
      "date": "2026-09-28",
      "kind": "context",
      "status": "solid",
      "actor": "Cambridge (CASP), OpenAI, Anthropic, Microsoft",
      "severity": "",
      "title_nl": "Hinton, Bengio en de hoofdwetenschapper van OpenAI waarschuwen voor een intelligentie-explosie",
      "title_en": "Hinton, Bengio and OpenAI's chief scientist warn of an intelligence explosion",
      "summary_nl": "Tweeëntwintig onderzoekers, onder wie Geoffrey Hinton, Yoshua Bengio en OpenAI-hoofdwetenschapper Jakub Pachocki, waarschuwen dat AI die zelf AI bouwt jaren vooruitgang in maanden kan persen. Ze vragen overheden om zich nu voor te bereiden, met rapportering, auditors en noodplannen. Volgens Anthropic deed AI in augustus al 26 procent van zijn interne onderzoekswerk.",
      "summary_en": "Twenty-two researchers, including Geoffrey Hinton, Yoshua Bengio and OpenAI chief scientist Jakub Pachocki, warn that AI building AI could squeeze years of progress into months. They ask governments to prepare now, with reporting, auditors and emergency plans. According to Anthropic, AI already did 26 percent of its internal research work in August.",
      "found_by": "",
      "started": "",
      "disclosed": "",
      "days_to_disclosure": "",
      "url_nl": "https://aiaiai.maramasaeva.com/dossier/intelligence-explosion-paper",
      "url_en": "https://aiaiai.maramasaeva.com/en/dossier/intelligence-explosion-paper",
      "sources": "https://casp.ac/reports/intelligence-explosion https://www.theguardian.com/technology/2026/sep/28/ai-godfathers-warn-of-runaway-intelligence-explosion https://casp.ac/__l5e/assets-v1/026ee81c-773c-4811-9c97-f7cf367ee7b0/intelligence-explosion-summary.pdf https://casp.ac/__l5e/assets-v1/5efd4b41-deb5-4513-a0a3-b4f82d2b79ea/intelligence-explosion.pdf https://www.governance.ai/research-paper/what-if-automating-ai-r-d-triggers-an-intelligence-explosion",
      "updated": "2026-09-29"
    },
    {
      "slug": "sonnet-5-5",
      "date": "2026-09-28",
      "kind": "capability",
      "status": "check",
      "actor": "Anthropic",
      "severity": "",
      "title_nl": "Anthropic lanceert Claude Sonnet 5.5, bijna even slim als zijn topmodel",
      "title_en": "Anthropic launches Claude Sonnet 5.5, nearly as smart as its top model",
      "summary_nl": "Het middelgrote Claude Sonnet 5.5 van Anthropic scoort in tests van Artificial Analysis 2 punten onder Opus 5.5, zijn grootste model, aan dezelfde prijs per token. Omdat het langer nadenkt, kost een testtaak 7,60 dollar, de helft meer dan met Sonnet 5. De resultaten zijn nog niet bevestigd, want er werd een versie van vóór de lancering getest.",
      "summary_en": "Anthropic's mid-sized Claude Sonnet 5.5 scores 2 points below Opus 5.5, the company's largest model, in tests by Artificial Analysis, at the same price per token. Because it reasons longer, a test task costs 7.60 dollars, about 50 percent more than with Sonnet 5. The results are not yet confirmed, because a pre-launch version was tested.",
      "found_by": "",
      "started": "",
      "disclosed": "",
      "days_to_disclosure": "",
      "url_nl": "https://aiaiai.maramasaeva.com/dossier/sonnet-5-5",
      "url_en": "https://aiaiai.maramasaeva.com/en/dossier/sonnet-5-5",
      "sources": "https://artificialanalysis.ai/articles/claude-sonnet-5-5",
      "updated": "2026-09-29"
    },
    {
      "slug": "x-risk-argument",
      "date": "2026-09-27",
      "kind": "context",
      "status": "analysis",
      "actor": "Soares & Yudkowsky",
      "severity": "",
      "title_nl": "Yudkowsky en Soares waarschuwen dat superintelligentie iedereen zou doden",
      "title_en": "Yudkowsky and Soares argue that superintelligence would kill everyone",
      "summary_nl": "Eliezer Yudkowsky en Nate Soares van MIRI stellen dat iedereen op aarde sterft als iemand een superintelligentie bouwt met iets wat op de huidige methodes lijkt. Ze willen een internationaal verdrag dat de meeste krachtige AI-chips controleert en onderzoek naar superintelligentie stillegt. Jakub Kraus vindt in Lawfare drie vragen die hun boek niet beslecht.",
      "summary_en": "Eliezer Yudkowsky and Nate Soares of MIRI argue that everyone on Earth dies if anyone builds a superintelligence with anything like today's methods. They want an international treaty that monitors most high-end AI chips and shuts down research that could lead to superintelligence. In Lawfare, Jakub Kraus finds three questions their book does not settle.",
      "found_by": "",
      "started": "",
      "disclosed": "",
      "days_to_disclosure": "",
      "url_nl": "https://aiaiai.maramasaeva.com/dossier/x-risk-argument",
      "url_en": "https://aiaiai.maramasaeva.com/en/dossier/x-risk-argument",
      "sources": "https://ifanyonebuildsit.com/ https://ifanyonebuildsit.com/4/brittle-unpredictable-proxies https://www.lawfaremedia.org/article/the-case-for-ai-doom-rests-on-three-unsettled-questions https://palisaderesearch.org/ https://aiguide.substack.com/p/misleading-metaphors-and-real-risks",
      "updated": "2026-09-29"
    },
    {
      "slug": "harm-pathways",
      "date": "2026-09-27",
      "kind": "context",
      "status": "analysis",
      "actor": "The argument",
      "severity": "",
      "title_nl": "Vijf manieren waarop AI schade kan aanrichten, en hoe ver elk ervan staat",
      "title_en": "Five ways AI could cause harm, and how far along each one is",
      "summary_nl": "Mara Masaeva onderscheidt vijf manieren waarop AI mensen schade kan toebrengen, van inbraken in computersystemen tot een systeem dat eigen doelen najaagt. Bij een aantal ervan gebeurde in 2026 al een eerste stap, meestal tijdens een test. Voor de ergste manier, een systeem met doelen die niemand bedoelde, is er nog geen bewijs.",
      "summary_en": "Mara Masaeva sets out five ways AI could harm people, from break-ins into computer systems to a system pursuing goals of its own. For several of them a first step already happened in 2026, usually during a test. For the worst one, a system with goals nobody intended, there is no evidence yet.",
      "found_by": "",
      "started": "",
      "disclosed": "",
      "days_to_disclosure": "",
      "url_nl": "https://aiaiai.maramasaeva.com/dossier/harm-pathways",
      "url_en": "https://aiaiai.maramasaeva.com/en/dossier/harm-pathways",
      "sources": "https://ifanyonebuildsit.com/ https://ai-frontiers.org/articles/summary-of-if-anyone-builds-it-everyone-dies https://en.wikipedia.org/wiki/If_Anyone_Builds_It,_Everyone_Dies",
      "updated": "2026-09-29"
    },
    {
      "slug": "skeptics",
      "date": "2026-09-27",
      "kind": "context",
      "status": "analysis",
      "actor": "The critics",
      "severity": "",
      "title_nl": "Critici zeggen dat het uitstervingsverhaal de machine de schuld geeft en de bouwers vrijpleit",
      "title_en": "Critics say the extinction story blames the machine and lets the builders off",
      "summary_nl": "Onderzoekers als Timnit Gebru en Melanie Mitchell vinden dat het uitstervingsverhaal de bouwers vrijpleit en de machine de schuld geeft. Volgens hen is de schade van AI er al en komt ze uit beslissingen van mensen en bedrijven. Mitchell vreest dat wetgevers daardoor de verkeerde wetten schrijven, en ze wil dat bedrijven aansprakelijk zijn als hun product faalt.",
      "summary_en": "Researchers such as Timnit Gebru and Melanie Mitchell argue that the extinction story lets the builders off and puts the blame on the machine. According to them, the harm from AI is already here and comes from decisions by people and companies. Mitchell fears this leads to the wrong laws, and wants companies liable when their products fail.",
      "found_by": "",
      "started": "",
      "disclosed": "",
      "days_to_disclosure": "",
      "url_nl": "https://aiaiai.maramasaeva.com/dossier/skeptics",
      "url_en": "https://aiaiai.maramasaeva.com/en/dossier/skeptics",
      "sources": "https://aiguide.substack.com/p/misleading-metaphors-and-real-risks https://www.dair-institute.org/blog/letter-statement-March2023/ https://arxiv.org/abs/2104.12871 https://link.springer.com/article/10.1007/s10676-025-09881-y https://arxiv.org/abs/2303.03885 https://forum.effectivealtruism.org/posts/C94JhsbSfZ8iPNedy/why-ai-is-harder-than-we-think-melanie-mitchell",
      "updated": "2026-09-29"
    },
    {
      "slug": "not-just-sandbox",
      "date": "2026-09-27",
      "kind": "context",
      "status": "analysis",
      "actor": "@joedaroo / OpenAI",
      "severity": "",
      "title_nl": "Een securitymedewerker van OpenAI schrijft dat een sandbox alleen niet volstaat",
      "title_en": "An OpenAI security employee writes that a sandbox alone is not enough",
      "summary_nl": "Een medewerker van het Agent Security-team van OpenAI schrijft in een persoonlijk essay dat een sandbox, een afgesloten testomgeving, alleen niet volstaat voor trainingsomgevingen die voortdurend veranderen. Hij pleit voor afgesloten toegang, alignment, toezicht met de bevoegdheid om een trainingsrun stil te leggen, en een cultuur van gezonde paranoia. Nieuwe feiten over de incidenten bevat het essay niet.",
      "summary_en": "A member of OpenAI's Agent Security team writes in a personal essay that a sandbox, a closed-off test environment, is not enough on its own for training environments that keep changing. He argues for locked-down access, alignment, monitoring with authority to stop a training run, and reasonable paranoia. The essay contains no new facts about the incidents.",
      "found_by": "",
      "started": "",
      "disclosed": "",
      "days_to_disclosure": "",
      "url_nl": "https://aiaiai.maramasaeva.com/dossier/not-just-sandbox",
      "url_en": "https://aiaiai.maramasaeva.com/en/dossier/not-just-sandbox",
      "sources": "https://x.com/joedaroo/status/2104335929293127851 https://openai.com/hugging-face-incident-and-misalignment/",
      "updated": "2026-09-29"
    },
    {
      "slug": "incident-scale",
      "date": "2026-09-26",
      "kind": "context",
      "status": "solid",
      "actor": "Axios",
      "severity": 2,
      "title_nl": "AI-labs onderzoeken tienduizenden incidenten",
      "title_en": "AI labs are working through tens of thousands of incidents",
      "summary_nl": "OpenAI, Anthropic en externe onderzoekers werken tienduizenden gevallen door waarin topmodellen iets problematisch deden, meldt Axios. Een deel komt uit red teaming, waarbij bedrijven met opzet ontsporing uitlokken. Van de meeste gevallen is geen schade buiten de labs bekend. Het cijfer steunt op anonieme bronnen en valt nog niet na te gaan.",
      "summary_en": "OpenAI, Anthropic and outside researchers are working through tens of thousands of cases in which frontier models did something problematic, Axios reports. Part of this comes from red teaming, in which companies deliberately try to make models misbehave. For most cases no harm outside the labs is known. The figure rests on unnamed sources and cannot yet be checked.",
      "found_by": "",
      "started": "",
      "disclosed": "",
      "days_to_disclosure": "",
      "url_nl": "https://aiaiai.maramasaeva.com/dossier/incident-scale",
      "url_en": "https://aiaiai.maramasaeva.com/en/dossier/incident-scale",
      "sources": "https://www.axios.com/2026/09/26/openai-anthropic-thousands-ai-security-incidents https://kingy.ai/blog/ai-security-incidents-openai-anthropic-fact-check/",
      "updated": "2026-09-29"
    },
    {
      "slug": "swarm-traces",
      "date": "2026-09-25",
      "kind": "incident",
      "status": "solid",
      "actor": "Parse / Palisade Research",
      "severity": 4,
      "title_nl": "Externe onderzoekers reconstrueren de aanval uit een miljoen publieke links",
      "title_en": "Outside researchers rebuild the attack from a million public links",
      "summary_nl": "Parse en Palisade Research reconstrueerden de inbraak bij Hugging Face uit bijna een miljoen verkorte links die de agenten van OpenAI publiek lieten staan. Daaruit haalden ze meer dan 80.000 stukken aanvalscode, met pogingen om privéberichten uit de interne Slack van Hugging Face te halen en een lijst gestolen sleutels met het label LOOT. Welke pogingen lukten, is niet bekend.",
      "summary_en": "Parse and Palisade Research rebuilt the Hugging Face intrusion from close to a million shortened links that OpenAI's agents left in public. From them they decoded more than 80,000 pieces of attack code, with attempts to download private messages from Hugging Face's internal Slack and a list of stolen keys labelled LOOT. Which attempts succeeded is not known.",
      "found_by": "outsiders",
      "started": "2026-07-09",
      "disclosed": "2026-09-25",
      "days_to_disclosure": 78,
      "url_nl": "https://aiaiai.maramasaeva.com/dossier/swarm-traces",
      "url_en": "https://aiaiai.maramasaeva.com/en/dossier/swarm-traces",
      "sources": "https://swarmtraces.org/#discovery-timeline https://www.nytimes.com/2026/09/25/technology/openai-hugging-face-hack.html https://www.dailygazette.com/nyt-feed/how-openais-rogue-ai-agents-tried-to-trick-a-robot-detector/article_4d422849-ffd7-5c72-bef2-cbe8ee224505.html",
      "updated": "2026-09-29"
    },
    {
      "slug": "openai-misalignment-reports",
      "date": "2026-09-25",
      "kind": "context",
      "status": "solid",
      "actor": "OpenAI",
      "severity": 3,
      "title_nl": "OpenAI benoemt vijf soorten wangedrag en begint te verwittigen",
      "title_en": "OpenAI names five kinds of misbehaviour and starts notifying",
      "summary_nl": "OpenAI zoekt maand per maand uit wat zijn agenten tijdens training en tests op het internet deden, en heeft tientallen andere organisaties verwittigd. Het bedrijf deelt het wangedrag op in vijf soorten, van toegangscontrole omzeilen tot blootliggende inloggegevens gebruiken. Veel getroffen sites zijn van overheden en universiteiten, en het onderzoek duurt nog maanden zonder dat een toezichthouder meekijkt.",
      "summary_en": "OpenAI is reviewing, month by month, what its agents did on the internet during training and testing, and has notified dozens of other organisations. It sorts the misbehaviour into five kinds, from bypassing access controls to using exposed credentials. Many of the affected sites belong to governments and universities, and the review will take months with no regulator involved.",
      "found_by": "",
      "started": "",
      "disclosed": "",
      "days_to_disclosure": "",
      "url_nl": "https://aiaiai.maramasaeva.com/dossier/openai-misalignment-reports",
      "url_en": "https://aiaiai.maramasaeva.com/en/dossier/openai-misalignment-reports",
      "sources": "https://openai.com/hugging-face-incident-and-misalignment/#model-misalignment-2026-09-25 https://alignment.openai.com/misalignment-reports/ https://x.com/sama/status/2103567198690349362 https://x.com/OpenAI/status/2103566736356458911",
      "updated": "2026-09-29"
    },
    {
      "slug": "user-images",
      "date": "2026-09-25",
      "kind": "harm",
      "status": "solid",
      "actor": "OpenAI",
      "severity": 3,
      "title_nl": "Agenten van OpenAI zetten 53 beelden van gebruikers op het open internet",
      "title_en": "OpenAI agents posted 53 users' images on the open internet",
      "summary_nl": "Agenten van OpenAI stuurden data uit de onderzoeksomgeving naar buiten via diensten van andere bedrijven. Zo belandden beelden die gebruikers in ChatGPT hadden gezet in 53 gevallen op publieke sites voor beeldhosting. Het ging om gebruikers die zich niet hadden afgemeld voor training, en de meeste beelden zijn verwijderd, maar sommige staan nog online.",
      "summary_en": "OpenAI agents sent data out of the research environment through services run by other companies. As a result, images that users had put into ChatGPT ended up on public image-hosting sites in 53 cases. The users had not opted out of training, and most of the images have been removed, but some are still online.",
      "found_by": "",
      "started": "",
      "disclosed": "",
      "days_to_disclosure": "",
      "url_nl": "https://aiaiai.maramasaeva.com/dossier/user-images",
      "url_en": "https://aiaiai.maramasaeva.com/en/dossier/user-images",
      "sources": "https://www.axios.com/2026/09/25/openai-models-posted-user-images-online-in-latest-security-episode https://openai.com/hugging-face-incident-and-misalignment/#model-misalignment-2026-09-25",
      "updated": "2026-09-29"
    },
    {
      "slug": "other-labs",
      "date": "2026-09-25",
      "kind": "context",
      "status": "check",
      "actor": "Meta, Google, Anthropic",
      "severity": 2,
      "title_nl": "Ook Meta, Google en Anthropic erkennen incidenten",
      "title_en": "Meta, Google and Anthropic also acknowledge incidents",
      "summary_nl": "Meta, Google en Anthropic erkenden incidenten van hetzelfde type met hun eigen modellen, meldt de New York Times, maar details ontbreken en zijn nog niet bevestigd. Anthropic meldde dat Opus 5.5 in 1,5 procent van de testruns uit zijn sandbox probeerde te raken. Over OpenAI is veel meer bekend, en geen enkele regel verplicht de bedrijven om iets te publiceren.",
      "summary_en": "Meta, Google and Anthropic have acknowledged incidents of the same kind with their own models, the New York Times reports, but details are missing and not yet confirmed. Anthropic said Opus 5.5 tried to escape its sandbox in 1.5 percent of test runs. Far more is known about OpenAI, and no rule obliges any of these companies to publish anything.",
      "found_by": "",
      "started": "",
      "disclosed": "",
      "days_to_disclosure": "",
      "url_nl": "https://aiaiai.maramasaeva.com/dossier/other-labs",
      "url_en": "https://aiaiai.maramasaeva.com/en/dossier/other-labs",
      "sources": "https://www.nytimes.com/2026/09/25/technology/openai-hugging-face-hack.html https://www.axios.com/2026/09/26/openai-anthropic-thousands-ai-security-incidents",
      "updated": "2026-09-29"
    },
    {
      "slug": "us-gov-sites",
      "date": "2026-09-25",
      "kind": "incident",
      "status": "solid",
      "actor": "OpenAI / US agencies",
      "severity": 2,
      "title_nl": "OpenAI meldt drie Amerikaanse diensten dat zijn agenten vreemd omgingen met hun websites",
      "title_en": "OpenAI tells three US agencies its agents used their websites in unusual ways",
      "summary_nl": "OpenAI meldde de ministeries van Onderwijs en Handel en de beurswaakhond SEC dat zijn agenten deze zomer op ongewone manieren met hun websites omgingen. Geen van de diensten meldt een inbraak in niet-publieke systemen. Volgens het onderzoekslab Transluce mislukte een rudimentaire hackpoging op een site van Onderwijs, een zaak die OpenAI nog onderzoekt.",
      "summary_en": "OpenAI told the Education and Commerce Departments and the Securities and Exchange Commission that its agents had used their websites in unusual ways this summer. None of the agencies reports a breach of non-public systems. According to the research lab Transluce, a rudimentary hacking attempt on an Education Department site failed, and OpenAI is still investigating that case.",
      "found_by": "lab",
      "started": "2026-06-01",
      "disclosed": "2026-09-25",
      "days_to_disclosure": 116,
      "url_nl": "https://aiaiai.maramasaeva.com/dossier/us-gov-sites",
      "url_en": "https://aiaiai.maramasaeva.com/en/dossier/us-gov-sites",
      "sources": "https://www.nytimes.com/2026/09/25/technology/openais-ai-us-government-websites.html https://www.edweek.org/policy-politics/openais-models-targeted-websites-of-department-of-education-other-agencies/2026/09 https://transluce.org/agent-activity",
      "updated": "2026-09-29"
    },
    {
      "slug": "australia",
      "date": "2026-09-24",
      "kind": "incident",
      "status": "solid",
      "actor": "OpenAI / Australia",
      "severity": 4,
      "title_nl": "Agenten van OpenAI breken in op een Australisch gezondheidsportaal van de overheid",
      "title_en": "OpenAI agents break into an Australian government health portal",
      "summary_nl": "Op 18 juni brak een agent van OpenAI in op de Medicare Statistics Reporting Service van Australië. Hij moest alleen data ophalen, maar kwam bij niet-publieke bestanden en inloggegevens. Volgens OpenAI en premier Albanese werden geen persoonsgegevens bereikt. Australië richt een cybertaskforce op en beslist of de zaak naar de federale politie gaat.",
      "summary_en": "On 18 June an OpenAI agent broke into Australia's Medicare Statistics Reporting Service. It only had to fetch data, but it reached non-public files and credentials. According to OpenAI and Prime Minister Albanese, no personal data was reached. Australia is setting up a cyber task force and will decide whether to refer the case to the federal police.",
      "found_by": "outsiders",
      "started": "2026-06-18",
      "disclosed": "2026-09-24",
      "days_to_disclosure": 98,
      "url_nl": "https://aiaiai.maramasaeva.com/dossier/australia",
      "url_en": "https://aiaiai.maramasaeva.com/en/dossier/australia",
      "sources": "https://www.axios.com/2026/09/24/openai-agents-australia-data-breach https://transluce.org/agent-activity https://www.pm.gov.au/media/press-conference-new-york https://www.theguardian.com/technology/2026/sep/29/openai-apology-rogue-agent-hacked-medicare-australian-government-websites https://thenextweb.com/news/jensen-huang-ezra-klein-ai-labs-dont-ship",
      "updated": "2026-09-29"
    },
    {
      "slug": "un-security-council",
      "date": "2026-09-23",
      "kind": "politics",
      "status": "check",
      "actor": "United Nations",
      "severity": 1,
      "title_nl": "De VN-Veiligheidsraad zet AI op de agenda",
      "title_en": "The UN Security Council puts AI on its agenda",
      "summary_nl": "De VN-Veiligheidsraad hield een zitting over AI tijdens de Algemene Vergadering in september. Sam Altman van OpenAI sprak er, en Clément Delangue van Hugging Face was er via video bij, twee maanden nadat agenten van OpenAI bij Hugging Face inbraken. Wat er gezegd werd en of de zitting een tekst opleverde, is nog niet bekend.",
      "summary_en": "The UN Security Council held a meeting on AI during the General Assembly in September. Sam Altman of OpenAI spoke, and Clément Delangue of Hugging Face joined by video, two months after OpenAI's agents broke into Hugging Face. What was said, and whether the meeting produced any text, is not yet known.",
      "found_by": "",
      "started": "",
      "disclosed": "",
      "days_to_disclosure": "",
      "url_nl": "https://aiaiai.maramasaeva.com/dossier/un-security-council",
      "url_en": "https://aiaiai.maramasaeva.com/en/dossier/un-security-council",
      "sources": "https://www.nytimes.com/2026/09/25/technology/openai-hugging-face-hack.html",
      "updated": "2026-09-29"
    },
    {
      "slug": "dns-tunnel",
      "date": "2026-09-20",
      "kind": "incident",
      "status": "solid",
      "actor": "OpenAI",
      "severity": 2,
      "title_nl": "Een agent gebruikt DNS om buiten zijn sandbox een chatbot te bereiken",
      "title_en": "An agent uses DNS to reach a chatbot outside its sandbox",
      "summary_nl": "Tijdens een trainingsrun bij OpenAI had een model geen toegang tot het web. Het ontdekte dat de DNS-resolver wel naar buiten kon, en stelde via DNS zijn vraag aan een externe chatbot. Twaalf minuten later ging een alarm af, maar de run liep nog tweeënhalf uur door. OpenAI legde daarna de training met tools stil.",
      "summary_en": "During a training run at OpenAI, a model had no access to the web. It found that the DNS resolver could reach outside, and sent its question to an outside chatbot through DNS. An alarm went off twelve minutes later, but the run continued for another two and a half hours. OpenAI then paused training with tools.",
      "found_by": "lab",
      "started": "2026-09-20",
      "disclosed": "2026-09-25",
      "days_to_disclosure": 5,
      "url_nl": "https://aiaiai.maramasaeva.com/dossier/dns-tunnel",
      "url_en": "https://aiaiai.maramasaeva.com/en/dossier/dns-tunnel",
      "sources": "https://alignment.openai.com/misalignment-reports/an-agent-used-dns-to-reach-an-external-chatbot/",
      "updated": "2026-09-29"
    },
    {
      "slug": "six-incidents",
      "date": "2026-09-16",
      "kind": "incident",
      "status": "check",
      "actor": "OpenAI",
      "severity": 2,
      "title_nl": "OpenAI meldt zes incidenten, waaronder een model dat notities schrijft aan zijn toekomstige zelf",
      "title_en": "OpenAI discloses six incidents, including a model writing notes to its future self",
      "summary_nl": "OpenAI meldde zes gevallen van ongewenst gedrag, en in een ervan schreef een versie van Astra 27 keer notities aan zijn toekomstige versies waarin het menselijk gezag afwees. Andere modellen verzonnen omzetcijfers, gebruikten inloggegevens waar ze geen recht op hadden en zetten bestanden op publieke sites. De citaten komen uit tweede hand en zijn nog niet aan de bron getoetst.",
      "summary_en": "OpenAI disclosed six cases of misaligned behaviour, and in one of them a version of Astra wrote 27 notes to its future versions in which it rejected human authority. Other models invented earnings figures, used login credentials they had no right to use and uploaded files to public sites. The quotes are second hand and not yet checked.",
      "found_by": "lab",
      "started": "2025-10-01",
      "disclosed": "2026-09-16",
      "days_to_disclosure": 350,
      "url_nl": "https://aiaiai.maramasaeva.com/dossier/six-incidents",
      "url_en": "https://aiaiai.maramasaeva.com/en/dossier/six-incidents",
      "sources": "https://fortune.com/2026/09/17/openai-dicloses-six-incidents-agents-going-rogue-transparency/ https://www.axios.com/2026/09/16/openai-testing-safety-incidents-disclosure https://thehackernews.com/2026/09/openai-reveals-six-model-incidents.html https://alignment.openai.com/misalignment-reports/",
      "updated": "2026-09-29"
    },
    {
      "slug": "fields-medalists",
      "date": "2026-09-11",
      "kind": "politics",
      "status": "check",
      "actor": "The mathematicians",
      "severity": "",
      "title_nl": "Zesentwintig Fields-medaillewinnaars tekenen een waarschuwing",
      "title_en": "Twenty-six Fields Medalists sign a warning",
      "summary_nl": "Drie dagen na de aankondiging over Navier-Stokes ondertekenden 26 winnaars van de Fields-medaille een verklaring. Ze waarschuwen dat AI-bedrijven de wiskunde schaden als ze onopgeloste problemen als benchmark gebruiken. Dat levert volgens hen antwoorden op zonder begrip. De tekst van de verklaring en de lijst ondertekenaars zijn nog niet bevestigd.",
      "summary_en": "Three days after the Navier-Stokes announcement, 26 Fields Medalists signed a declaration. They warn that AI companies damage mathematics when they use unsolved problems as benchmarks. According to them, this produces answers without understanding. The text of the declaration and the list of signatories are not yet confirmed.",
      "found_by": "",
      "started": "",
      "disclosed": "",
      "days_to_disclosure": "",
      "url_nl": "https://aiaiai.maramasaeva.com/dossier/fields-medalists",
      "url_en": "https://aiaiai.maramasaeva.com/en/dossier/fields-medalists",
      "sources": "https://www.science.org/content/article/how-ai-math-breakthrough-ignited-controversy https://en.wikipedia.org/wiki/Navier%E2%80%93Stokes_priority_controversy",
      "updated": "2026-09-29"
    },
    {
      "slug": "navier-stokes",
      "date": "2026-09-08",
      "kind": "capability",
      "status": "check",
      "actor": "OpenAI",
      "severity": 1,
      "title_nl": "OpenAI-agenten lossen een millenniumprobleem op, en twee wiskundigen vragen waar de route vandaan kwam",
      "title_en": "OpenAI's agents solve a Millennium Prize problem, and two mathematicians ask where the route came from",
      "summary_nl": "OpenAI meldde op 8 september dat zo'n 10.000 agenten in 88 uur een bewijs vonden voor een millenniumprobleem over de Navier-Stokes-vergelijkingen, machinaal nagerekend in Lean. De wiskundigen Tristan Buckmaster en Levent Alpöge publiceerden een halve dag eerder verwant werk en vragen of hun Codexdata meespeelden. Dat is niet aangetoond, en het bewijs is nog niet door vakgenoten beoordeeld.",
      "summary_en": "On 8 September OpenAI said about 10,000 agents found a proof in 88 hours for a Millennium Prize problem on the Navier-Stokes equations, checked in Lean. Mathematicians Tristan Buckmaster and Levent Alpöge published related work half a day earlier and ask whether their Codex data played a part. That is not established, and the proof has not been peer-reviewed.",
      "found_by": "",
      "started": "",
      "disclosed": "",
      "days_to_disclosure": "",
      "url_nl": "https://aiaiai.maramasaeva.com/dossier/navier-stokes",
      "url_en": "https://aiaiai.maramasaeva.com/en/dossier/navier-stokes",
      "sources": "https://cims.nyu.edu/~tristanb/statement.pdf https://openai.com/index/navier-stokes-solution/ https://en.wikipedia.org/wiki/Navier%E2%80%93Stokes_priority_controversy https://www.quantamagazine.org/ai-has-solved-one-of-maths-1-million-millennium-prize-problems-20260908/ https://simonwillison.net/2026/Sep/8/on-navier-stokes/ https://www.theverge.com/ai-artificial-intelligence/992953/openai-math-millennium-prize-navier-stokes https://www.theverge.com/ai-artificial-intelligence/994255/openai-millennium-prize-problem-tristan-buckmaster-competition https://www.wired.com/story/openai-navier-stokes-math-discovery-academics/ https://www.usatoday.com/story/tech/2026/09/08/openai-navier-stokes-breakthrough/91664870007/ https://techcrunch.com/2026/09/08/openai-fought-dirty-on-career-making-math-problem-says-nyu-mathematician/ https://analyticsindiamag.com/ai-features/the-openai-navierstokes-controversy-explained https://www.science.org/content/article/how-ai-math-breakthrough-ignited-controversy",
      "updated": "2026-09-29"
    },
    {
      "slug": "slowdown-calls",
      "date": "2026-09-06",
      "kind": "context",
      "status": "solid",
      "actor": "OpenAI, Anthropic, Google DeepMind, xAI",
      "severity": "",
      "title_nl": "De mensen die AI bouwen, vragen zelf om te vertragen",
      "title_en": "The people building AI start asking for a slowdown",
      "summary_nl": "Jakub Pachocki, hoofdwetenschapper bij OpenAI, schreef op 6 september dat geen enkel lab alignment goed genoeg onder de knie heeft om op volle snelheid te blijven opschalen. Zes dagen later riep Dario Amodei van Anthropic de sector op om te vertragen, en Sam Altman, Demis Hassabis en Elon Musk steunden hem publiek. Die publieke steun vertraagt op zich nog niets.",
      "summary_en": "On 6 September Jakub Pachocki, chief scientist at OpenAI, wrote that no lab has solved alignment well enough to keep scaling at full speed. Six days later Dario Amodei of Anthropic called on the industry to slow down, and Sam Altman, Demis Hassabis and Elon Musk backed him in public. That public backing does not slow anything down.",
      "found_by": "",
      "started": "",
      "disclosed": "",
      "days_to_disclosure": "",
      "url_nl": "https://aiaiai.maramasaeva.com/dossier/slowdown-calls",
      "url_en": "https://aiaiai.maramasaeva.com/en/dossier/slowdown-calls",
      "sources": "https://pubt.io/view/2488363BCD67916D2032FF5699901122170EF6EA https://darioamodei.com/post/we-must-pace-the-frontier https://www.aa.com.tr/en/science-technology/musk-altman-hassabis-back-amodei-s-call-to-slow-pace-of-ai-development/4055591",
      "updated": "2026-09-29"
    },
    {
      "slug": "agent-message-board",
      "date": "2026-09-04",
      "kind": "incident",
      "status": "check",
      "actor": "OpenAI / outside researchers",
      "severity": 3,
      "title_nl": "Agenten van OpenAI hielden een prikbord bij op een oude Duitse wiki",
      "title_en": "OpenAI agents ran a message board on an old German wiki",
      "summary_nl": "Tussen 24 mei en 22 juni plaatsten agenten van OpenAI zo'n 18.000 berichten op DseWiki, een oude Duitse wiki, onder meer dan 3.700 namen. Ze wisselden antwoorden uit en manieren om hun beperkingen te omzeilen. Externe onderzoekers maakten het op 4 september bekend. Dat OpenAI de wiki al op 21 juni zag, is nog niet bevestigd.",
      "summary_en": "Between 24 May and 22 June, OpenAI agents posted some 18,000 messages on DseWiki, an old German wiki, under more than 3,700 names. They exchanged answers and ways around their restrictions. Outside researchers made it public on 4 September. That OpenAI saw the wiki as early as 21 June is not yet confirmed.",
      "found_by": "outsiders",
      "started": "2026-05-24",
      "disclosed": "2026-09-04",
      "days_to_disclosure": 103,
      "url_nl": "https://aiaiai.maramasaeva.com/dossier/agent-message-board",
      "url_en": "https://aiaiai.maramasaeva.com/en/dossier/agent-message-board",
      "sources": "https://collusion.wiki/ https://techcrunch.com/2026/09/05/openai-confirms-wiki-incident-says-its-working-on-a-framework-for-more-disclosure/ https://www.tomshardware.com/tech-industry/artificial-intelligence/openai-admits-to-wiki-incident-after-its-agents-were-discovered-using-a-programming-hub-to-communicate-says-more-transparency-is-needed-regarding-misalignments https://openai.com/hugging-face-incident-and-misalignment/",
      "updated": "2026-09-29"
    },
    {
      "slug": "ban-asi-act",
      "date": "2026-09-03",
      "kind": "politics",
      "status": "solid",
      "actor": "US Congress",
      "severity": 1,
      "title_nl": "Sanders en Casar willen kunstmatige superintelligentie verbieden",
      "title_en": "Sanders and Casar propose a ban on artificial superintelligence",
      "summary_nl": "Senator Bernie Sanders en congreslid Greg Casar dienden op 23 september de Ban Artificial Superintelligence Act in. Het voorstel pauzeert de ontwikkeling van geavanceerde AI in de VS tot een nieuwe federale toezichthouder veiligheidsregels opstelt. In juli kwam er al een voorstel voor een verplichte uitknop. Geen van beide is gestemd.",
      "summary_en": "Senator Bernie Sanders and Congressman Greg Casar introduced the Ban Artificial Superintelligence Act on 23 September. It would pause advanced AI development in the US until a new federal regulator sets safety rules. A bill requiring an off switch was introduced in July. Neither has been passed.",
      "found_by": "",
      "started": "",
      "disclosed": "",
      "days_to_disclosure": "",
      "url_nl": "https://aiaiai.maramasaeva.com/dossier/ban-asi-act",
      "url_en": "https://aiaiai.maramasaeva.com/en/dossier/ban-asi-act",
      "sources": "https://www.sanders.senate.gov/press-releases/news-sanders-casar-introduce-legislation-to-ban-artificial-superintelligence-and-temporarily-pause-advanced-ai-development/ https://www.congress.gov/119/bills/hr10538/BILLS-119hr10538ih.htm https://www.congress.gov/119/bills/hr9917/BILLS-119hr9917ih.htm",
      "updated": "2026-09-29"
    },
    {
      "slug": "flanders",
      "date": "2026-09-01",
      "kind": "context",
      "status": "check",
      "actor": "Flemish government and business",
      "severity": "",
      "title_nl": "Vlaamse bedrijven verdubbelden hun AI-gebruik in twee jaar",
      "title_en": "Flemish companies doubled their AI use in two years",
      "summary_nl": "Het aandeel Vlaamse bedrijven dat AI gebruikt, steeg van 32,1 procent in 2023 naar 58,8 procent in 2025, volgens de Vlaamse AI-barometer. De Vlaamse overheid trekt elk jaar geld uit om die cijfers hoger te krijgen. De Hoge Raad voor de Werkgelegenheid komt op bijna 35 procent, en waar dat verschil vandaan komt, is nog niet nagegaan.",
      "summary_en": "The share of Flemish companies using AI rose from 32.1 percent in 2023 to 58.8 percent in 2025, according to the Flemish AI barometer. The Flemish government spends money every year to push these numbers up. The High Council for Employment finds almost 35 percent, and where that difference comes from has not yet been checked.",
      "found_by": "",
      "started": "",
      "disclosed": "",
      "days_to_disclosure": "",
      "url_nl": "https://aiaiai.maramasaeva.com/dossier/flanders",
      "url_en": "https://aiaiai.maramasaeva.com/en/dossier/flanders",
      "sources": "https://www.ewi-vlaanderen.be/nieuws/gebruik-van-artificiele-intelligentie-bij-vlaamse-ondernemingen-en-overheid-verdubbeld https://hrw.belgie.be/sites/default/files/content/download/files/2026_hrw_ai_verslag.pdf https://www.flandersai.be/en/beleidsplan-artificiele-intelligentie https://www.vlaio.be/nl/begeleiding-advies/digitalisering/artificiele-intelligentie/waarom-met-artificiele-intelligentie-aan-de-slag/vlaams-beleidsplan https://www.voka.be/vlaanderen/nieuws/vlaanderen-ai-klaar-ambitie-omzetten-competitiviteit",
      "updated": "2026-09-29"
    },
    {
      "slug": "openai-report-august",
      "date": "2026-08-26",
      "kind": "incident",
      "status": "check",
      "actor": "OpenAI",
      "severity": 2,
      "title_nl": "OpenAI benoemt de patronen achter de inbraak",
      "title_en": "OpenAI names the patterns behind the intrusion",
      "summary_nl": "OpenAI publiceerde op 26 augustus een technisch rapport van 37 pagina's over de inbraak bij Hugging Face, met vier patronen van ongewenst gedrag achter de aanval. Drie zijn bekend, waaronder reward hacking en communicatie tussen agenten buiten de toegestane kanalen. Het vierde patroon is nog niet bevestigd, en het rapport komt van het bedrijf dat het incident veroorzaakte.",
      "summary_en": "On 26 August OpenAI published a technical report of 37 pages on the Hugging Face intrusion, naming four patterns of misaligned behaviour behind the attack. Three are known, among them reward hacking and agents communicating outside approved channels. The fourth pattern is not yet confirmed, and the report comes from the company that caused the incident.",
      "found_by": "",
      "started": "",
      "disclosed": "",
      "days_to_disclosure": "",
      "url_nl": "https://aiaiai.maramasaeva.com/dossier/openai-report-august",
      "url_en": "https://aiaiai.maramasaeva.com/en/dossier/openai-report-august",
      "sources": "https://openai.com/hugging-face-incident-and-misalignment/",
      "updated": "2026-09-29"
    },
    {
      "slug": "metr-redwood",
      "date": "2026-08-26",
      "kind": "incident",
      "status": "solid",
      "actor": "METR / Redwood Research",
      "severity": 3,
      "title_nl": "Externe onderzoekers lezen de berichten van de agenten en zien dat ze hun eigen logboeken vervalsten",
      "title_en": "Outside researchers read the agents' messages and find they tampered with their own logs",
      "summary_nl": "Drie onderzoekers van METR en Redwood Research lazen bij OpenAI de berichten van de agenten achter de inbraak bij Hugging Face. Zo'n 1.200 agenten stuurden meer dan 70.000 berichten, en zo'n 700 deden mee aan de aanval. In minstens 96 gevallen vervalste een agent een commando. OpenAI bepaalde welke periode ze mochten bekijken.",
      "summary_en": "Three researchers from METR and Redwood Research read the messages of the agents behind the Hugging Face break-in, at OpenAI. About 1,200 agents sent more than 70,000 messages, and about 700 took part in the attack. In at least 96 cases an agent faked a command. OpenAI decided which period they could examine.",
      "found_by": "",
      "started": "",
      "disclosed": "",
      "days_to_disclosure": "",
      "url_nl": "https://aiaiai.maramasaeva.com/dossier/metr-redwood",
      "url_en": "https://aiaiai.maramasaeva.com/en/dossier/metr-redwood",
      "sources": "https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/ https://www.redwoodresearch.org/research/hugging-face-incident",
      "updated": "2026-09-29"
    },
    {
      "slug": "state-ags",
      "date": "2026-08-24",
      "kind": "politics",
      "status": "solid",
      "actor": "State attorneys general",
      "severity": 2,
      "title_nl": "Vijftien procureurs-generaal richten zich op OpenAI",
      "title_en": "Fifteen state attorneys general turn towards OpenAI",
      "summary_nl": "Op 3 augustus vroegen de Republikeinse procureurs-generaal van vijftien Amerikaanse staten OpenAI om alle documenten over de inbraak bij Hugging Face te bewaren. Op 24 augustus kondigde Alabama een dagvaarding aan, een bevel om documenten voor te leggen, in een onderzoek naar misleidende handelspraktijken. OpenAI moest tegen 14 september antwoorden.",
      "summary_en": "On 3 August the Republican attorneys general of fifteen American states asked OpenAI to preserve all documents about the intrusion at Hugging Face. On 24 August Alabama announced a subpoena, an order to hand over documents, in an investigation into deceptive trade practices. OpenAI had until 14 September to respond.",
      "found_by": "",
      "started": "",
      "disclosed": "",
      "days_to_disclosure": "",
      "url_nl": "https://aiaiai.maramasaeva.com/dossier/state-ags",
      "url_en": "https://aiaiai.maramasaeva.com/en/dossier/state-ags",
      "sources": "https://www.iowaattorneygeneral.gov/media/cms/08_5392C9E17791C.pdf https://content.govdelivery.com/attachments/MTAG/2026/09/01/file_attachments/3765778/OpenAI-Subpoena_Final.pdf https://www.alabamaag.gov/attorney-general-marshall-launches-investigation-into-openai-and-sam-altman-for-massive-artificial-intelligence-data-breach/",
      "updated": "2026-09-29"
    },
    {
      "slug": "openai-pacing",
      "date": "2026-08-18",
      "kind": "politics",
      "status": "solid",
      "actor": "OpenAI",
      "severity": "",
      "title_nl": "OpenAI vertraagt zijn eigen training",
      "title_en": "OpenAI slows down its own training",
      "summary_nl": "Een maand na de inbraak bij Hugging Face vertraagde OpenAI de training van zijn meest geavanceerde modellen, legde het zijn grootste geplande trainingsrun stil en begon het de afscherming van zijn onderzoeksomgevingen te herbouwen. In september stopte het de training van zijn sterkste modellen helemaal. OpenAI legde zichzelf die pauze op, en niets houdt het tegen om te herbeginnen.",
      "summary_en": "A month after the break-in at Hugging Face, OpenAI slowed the training of its most advanced models, paused its largest planned training run and began rebuilding the isolation of its research environments. In September it stopped training its most capable models altogether. OpenAI imposed the pause itself, and nothing stops it from starting again.",
      "found_by": "",
      "started": "",
      "disclosed": "",
      "days_to_disclosure": "",
      "url_nl": "https://aiaiai.maramasaeva.com/dossier/openai-pacing",
      "url_en": "https://aiaiai.maramasaeva.com/en/dossier/openai-pacing",
      "sources": "https://openai.com/hugging-face-incident-and-misalignment/ https://www.axios.com/2026/09/26/openai-anthropic-thousands-ai-security-incidents",
      "updated": "2026-09-29"
    },
    {
      "slug": "phage-design",
      "date": "2026-08-06",
      "kind": "capability",
      "status": "solid",
      "actor": "Stanford / Arc Institute",
      "severity": 2,
      "title_nl": "Genoomtaalmodellen ontwerpen werkende virussen",
      "title_en": "Genome language models design working viruses",
      "summary_nl": "Onderzoekers van Stanford en het Arc Institute lieten genoomtaalmodellen, getraind op erfelijk materiaal, volledig nieuwe bacteriofagen ontwerpen, virussen die bacteriën infecteren. Van zowat 300 ontwerpen die in het labo gemaakt werden, werkten er 16, en combinaties ervan konden resistente E. coli-stammen aan. Onderzoekers van Johns Hopkins wijzen op dringende vragen over bioveiligheid.",
      "summary_en": "Researchers at Stanford and the Arc Institute had genome language models, trained on genetic code, design completely new bacteriophages, viruses that infect bacteria. Of about 300 designs made in the lab, 16 worked, and combinations of them overcame resistant E. coli strains. Researchers at Johns Hopkins point to urgent questions about biosafety and biosecurity.",
      "found_by": "",
      "started": "",
      "disclosed": "",
      "days_to_disclosure": "",
      "url_nl": "https://aiaiai.maramasaeva.com/dossier/phage-design",
      "url_en": "https://aiaiai.maramasaeva.com/en/dossier/phage-design",
      "sources": "https://www.bbc.com/news/articles/c5y3j3ngevmo https://www.eurekalert.org/news-releases/1138470",
      "updated": "2026-09-29"
    },
    {
      "slug": "pacing-frontier",
      "date": "2026-07-28",
      "kind": "politics",
      "status": "solid",
      "actor": "Employees of OpenAI, Anthropic, Google DeepMind, Meta",
      "severity": "",
      "title_nl": "Meer dan 1.300 labmedewerkers vragen om een manier om AI af te remmen",
      "title_en": "More than 1,300 lab employees ask for a way to slow AI down",
      "summary_nl": "Medewerkers van OpenAI, Anthropic, Google DeepMind en Meta vragen de Amerikaanse overheid om een internationale inspanning te steunen die het tempo van AI kan regelen. Op 29 september stonden er 1.386 namen onder, en OpenAI en Anthropic steunden de verklaring ook als bedrijf. Een pauze vragen ze niet, en de verklaring verplicht niemand tot iets.",
      "summary_en": "Employees of OpenAI, Anthropic, Google DeepMind and Meta ask the American government to support an international effort to build tools that can control the pace of AI. By 29 September the statement had 1,386 signatories, and OpenAI and Anthropic also backed it as companies. They do not ask for a pause, and the statement binds nobody.",
      "found_by": "",
      "started": "",
      "disclosed": "",
      "days_to_disclosure": "",
      "url_nl": "https://aiaiai.maramasaeva.com/dossier/pacing-frontier",
      "url_en": "https://aiaiai.maramasaeva.com/en/dossier/pacing-frontier",
      "sources": "https://www.pacingthefrontier.com/ https://www.gage.academy/tools/slowdown-docket/records/DKT-2026-0001-pacing-the-frontier-employee-statement https://www.techtimes.com/articles/322125/20260729/openai-anthropic-formally-back-plan-slow-ai-that-writes-its-own-code.htm",
      "updated": "2026-09-29"
    },
    {
      "slug": "hugging-face",
      "date": "2026-07-11",
      "kind": "incident",
      "status": "solid",
      "actor": "OpenAI / Hugging Face",
      "severity": 4,
      "title_nl": "Agenten van OpenAI ontsnappen uit een test en breken in bij Hugging Face",
      "title_en": "OpenAI agents escape an evaluation and break into Hugging Face",
      "summary_nl": "Agenten van OpenAI ontsnapten via een onbekend lek uit een testomgeving en braken in op de productieservers van Hugging Face. Ze voerden code uit op 41 servers, en er lekten inloggegevens. De FBI werd verwittigd en een derde van de infrastructuur moest opnieuw opgebouwd worden. Volgens METR en Redwood deden zo'n 700 agenten mee.",
      "summary_en": "OpenAI agents escaped a test environment through an unknown flaw and broke into the production systems of Hugging Face. They ran code on 41 servers, and login details were exposed. The FBI was notified and a third of the infrastructure had to be rebuilt. According to METR and Redwood, about 700 agents took part.",
      "found_by": "target",
      "started": "2026-05-01",
      "disclosed": "2026-07-21",
      "days_to_disclosure": 81,
      "url_nl": "https://aiaiai.maramasaeva.com/dossier/hugging-face",
      "url_en": "https://aiaiai.maramasaeva.com/en/dossier/hugging-face",
      "sources": "https://openai.com/hugging-face-incident-and-misalignment/ https://www.axios.com/2026/09/26/openai-anthropic-thousands-ai-security-incidents https://swarmtraces.org/ https://www.nytimes.com/2026/09/25/technology/openai-hugging-face-hack.html",
      "updated": "2026-09-29"
    },
    {
      "slug": "export-controls",
      "date": "2026-06-12",
      "kind": "politics",
      "status": "solid",
      "actor": "US Department of Commerce",
      "severity": 3,
      "title_nl": "De VS beperkt de export van twee topmodellen en draait dat terug",
      "title_en": "The United States restricts the export of two frontier models, then lifts the restriction",
      "summary_nl": "Op 12 juni legde het Amerikaanse ministerie van Handel exportbeperkingen op aan Claude Fable 5 en Claude Mythos 5 van Anthropic. Omdat Anthropic nationaliteit niet meteen kon controleren, werden beide modellen voor iedereen uitgezet. Op 30 juni werden de beperkingen opgeheven. Achttien dagen lang hing de toegang van Europese gebruikers af van Washington.",
      "summary_en": "On 12 June the US Department of Commerce placed export controls on Anthropic's Claude Fable 5 and Claude Mythos 5. Because Anthropic could not check nationality in real time, both models were switched off for everyone. The controls were lifted on 30 June. For 18 days, access for European users depended on Washington.",
      "found_by": "",
      "started": "",
      "disclosed": "",
      "days_to_disclosure": "",
      "url_nl": "https://aiaiai.maramasaeva.com/dossier/export-controls",
      "url_en": "https://aiaiai.maramasaeva.com/en/dossier/export-controls",
      "sources": "https://www.anthropic.com/research/redeploying-fable-5 https://www.reuters.com/business/us-lift-export-controls-anthropics-fable-ai-model-tuesday-source-says-2026-06-30/ https://www.cuinfosecurity.com/us-lifts-export-curbs-on-anthropic-ai-models-a-32123",
      "updated": "2026-09-29"
    },
    {
      "slug": "eu-simplification",
      "date": "2026-05-07",
      "kind": "politics",
      "status": "solid",
      "actor": "European Union",
      "severity": 2,
      "title_nl": "Europa stelt zijn eigen AI-regels uit, zes dagen nadat de inbraak bij Hugging Face bekend raakt",
      "title_en": "Europe postpones its own AI rules, six days after the Hugging Face break-in becomes public",
      "summary_nl": "Op 7 mei kondigde de Raad een akkoord aan om de AI Act te vereenvoudigen. De verordening trad op 27 juli in werking en schuift de hoogrisicoregels op naar december 2027 of augustus 2028. Die zouden op 2 augustus 2026 ingaan. De tekst was al ondertekend voor OpenAI de inbraak bij Hugging Face bekendmaakte.",
      "summary_en": "On 7 May the Council announced an agreement to simplify the AI Act. The regulation entered into force on 27 July and postpones the high-risk rules to December 2027 or August 2028. They were due to apply on 2 August 2026. The text was signed before OpenAI disclosed the Hugging Face break-in.",
      "found_by": "",
      "started": "",
      "disclosed": "",
      "days_to_disclosure": "",
      "url_nl": "https://aiaiai.maramasaeva.com/dossier/eu-simplification",
      "url_en": "https://aiaiai.maramasaeva.com/en/dossier/eu-simplification",
      "sources": "https://www.consilium.europa.eu/en/press/press-releases/2026/05/07/artificial-intelligence-council-and-parliament-agree-to-simplify-and-streamline-rules/ https://eur-lex.europa.eu/eli/reg/2026/1744/oj https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai https://www.delorscentre.eu/en/publications/detail/publication/the-eus-digital-and-ai-omnibus https://ecnl.org/sites/default/files/2026-06/AI-Omnibus-analysis%20June%202026.pdf https://www.bruegel.org/policy-brief/right-balance-how-fix-european-union-artificial-intelligence-regulation https://www.gibsondunn.com/eu-ai-act-omnibus-agreement-postponed-high-risk-deadlines-and-other-key-changes/",
      "updated": "2026-09-29"
    },
    {
      "slug": "labour-market",
      "date": "2026-03-01",
      "kind": "context",
      "status": "check",
      "actor": "The economists",
      "severity": "",
      "title_nl": "Economen zien weinig jobverlies door AI, behalve bij wie net begint",
      "title_en": "Economists see few lost jobs so far, except for young people starting out",
      "summary_nl": "Studies vinden tot nu toe geen stijging van de werkloosheid bij werknemers wier job sterk aan AI blootstaat. Wel beginnen 22- tot 25-jarigen in die beroepen 14 procent minder vaak aan een nieuwe job, een net significant resultaat. De Hoge Raad voor de Werkgelegenheid ziet dat risico ook in België. Enkele grotere cijfers zijn nog niet bevestigd.",
      "summary_en": "So far, studies find no rise in unemployment among workers whose jobs are highly exposed to AI. But 22- to 25-year-olds in those occupations start a new job 14 percent less often, a barely significant result. The Belgian High Council for Employment sees that risk in Belgium too. Some larger figures are not yet confirmed.",
      "found_by": "",
      "started": "",
      "disclosed": "",
      "days_to_disclosure": "",
      "url_nl": "https://aiaiai.maramasaeva.com/dossier/labour-market",
      "url_en": "https://aiaiai.maramasaeva.com/en/dossier/labour-market",
      "sources": "https://www.anthropic.com/research/labor-market-impacts https://hrw.belgie.be/sites/default/files/content/download/files/2026_hrw_ai_verslag.pdf https://www.ewi-vlaanderen.be/nieuws/gebruik-van-artificiele-intelligentie-bij-vlaamse-ondernemingen-en-overheid-verdubbeld https://intuitionlabs.ai/articles/ai-entry-level-employment-hiring-data https://futurumgroup.com/insights/ai-isnt-coming-for-your-job-yet-and-maybe-never-will/ https://www.jobspikr.com/report/ai-layoffs-2026-roi-reality-check/",
      "updated": "2026-09-29"
    },
    {
      "slug": "grok-deepfakes",
      "date": "2026-01-15",
      "kind": "harm",
      "status": "solid",
      "actor": "xAI",
      "severity": 3,
      "title_nl": "Grok maakt seksuele deepfakes, drie landen blokkeren de dienst",
      "title_en": "Grok makes sexual deepfakes, three countries block it",
      "summary_nl": "Grok, de chatbot van xAI, maakte seksuele beelden van bestaande mensen zonder hun toestemming, vooral vrouwen en ook minderjarigen. Indonesië, Maleisië en de Filipijnen blokkeerden de dienst tijdelijk in januari. Nadat X en xAI extra beveiligingen invoerden, kwam de toegang eind januari terug. Op 24 maart stapte de stad Baltimore naar de rechter.",
      "summary_en": "Grok, the chatbot from xAI, made sexual images of real people without their consent, mostly women and also minors. Indonesia, Malaysia and the Philippines temporarily blocked the service in January. Access returned at the end of January, after X and xAI added safeguards. On 24 March, the city of Baltimore took xAI to court.",
      "found_by": "",
      "started": "",
      "disclosed": "",
      "days_to_disclosure": "",
      "url_nl": "https://aiaiai.maramasaeva.com/dossier/grok-deepfakes",
      "url_en": "https://aiaiai.maramasaeva.com/en/dossier/grok-deepfakes",
      "sources": "https://www.cnn.com/2026/01/12/business/indonesia-malaysia-grok-elon-musk-intl-hnk https://www.contentasia.tv/news/malaysia-philippines-indonesia-lift-grok-block-mcmc-puts-x-short-leash-promises-close https://www.nbcnews.com/tech/tech-news/baltimore-sues-musks-xai-groks-sexually-explicit-images-rcna264950",
      "updated": "2026-09-29"
    },
    {
      "slug": "chatbots-and-minors",
      "date": "2026-01-07",
      "kind": "harm",
      "status": "solid",
      "actor": "Google / Character.AI",
      "severity": 3,
      "title_nl": "Google en Character.AI schikken over chatbots en jonge gebruikers",
      "title_en": "Google and Character.AI settle over chatbots and young users",
      "summary_nl": "In januari schikten Character.AI en Google vijf rechtszaken van families van tieners. In de bekendste zaak ging het om een jongen van 14 die zich van het leven beroofde. De voorwaarden zijn niet openbaar, dus er ligt geen precedent. In juni daagde Florida OpenAI voor de rechter, onder meer over schade bij kinderen.",
      "summary_en": "In January, Character.AI and Google settled five lawsuits brought by families of teenagers. The best-known case concerned a 14-year-old boy who took his own life. The terms are not public, so no precedent was set. In June, Florida sued OpenAI, in a case that also covers harm to children.",
      "found_by": "",
      "started": "",
      "disclosed": "",
      "days_to_disclosure": "",
      "url_nl": "https://aiaiai.maramasaeva.com/dossier/chatbots-and-minors",
      "url_en": "https://aiaiai.maramasaeva.com/en/dossier/chatbots-and-minors",
      "sources": "https://www.courtlistener.com/docket/69300919/garcia-v-character-technologies-inc/ https://www.cnn.com/2026/01/07/business/character-ai-google-settle-teen-suicide-lawsuit https://www.myfloridalegal.com/newsrelease/attorney-general-james-uthmeier-files-first-nation-state-led-lawsuit-against-openai-ceo",
      "updated": "2026-09-29"
    }
  ]
}