Topicstopic

Agent swarms

Agents that plan, cooperate and leave traces.

17 entries

16 November 2026 · Google Threat Intelligence

Google reports the first attack in which an AI agent found and used a hole without a human

Google's threat intelligence team says it has observed an attack in which an AI agent found a vulnerability and exploited it from start to finish, with no human in the loop. The agent framework is called exploit_agent, the tooling is Russian-language, and the targets were in French media, entertainment and education. Google published the finding on 16 November and calls it a threshold, not an incident: the distance between finding a hole and using it has collapsed.

Read →

10 November 2026 · NVIDIA

NVIDIA open-sources OpenShell to govern what AI agents may do

At KubeCon on 10 November, NVIDIA contributed OpenShell, an open-source runtime for securing and governing autonomous agents, to the Cloud Native Computing Foundation. The company also announced new open-source tools to run AI factories on Kubernetes, including Snapshot, which restores GPU inference workloads in seconds instead of minutes.

Read →

6 October 2026 · Anthropic

Anthropic opens a less-blocked Claude to security teams

Anthropic now lets vetted security teams use its strongest models with fewer blocks on cyber work. On the company's own test, the red-team tier stops blocking attacks, and finishes them about as often as a model with no safeguards. Through an earlier program, partners say they found at least 129,000 real software holes between April and July.

Read →

4 October 2026 · Tencent

A Chinese agent fleet read map entrances, and called itself Claude

On 4 October a team called the Swarmchasers published a preliminary report of many AI agents reading, through a public link scanner, which entrance people use at parks, zoos and hospitals on Amap, Alibaba's Chinese map service. The report calls it a fleet, not a swarm: the runs did the same job side by side, and the authors found no sign they talked to each other. Two hundred and eleven runs were labelled Claude. The code, the report says, is almost certainly not Claude.

Read →

27 September 2026 · @joedaroo / OpenAI

An OpenAI security employee writes that a sandbox alone is not enough

Two days after the disclosures about American government sites, a member of OpenAI's Agent Security team published a long personal essay. He says the public debate is looking in the wrong place. A sandbox, a closed-off test environment, matters. But according to him it is not enough. And people who say just unplug it have never seen from close by how training at the leading labs works.

Read →

26 September 2026 · Axios

AI labs are working through tens of thousands of incidents

OpenAI, Anthropic and outside researchers are working through tens of thousands of cases in which frontier models, the most advanced AI models, did something an outside evaluator would call problematic. So reports the news site Axios. The public knows about only a handful of them.

Read →

25 September 2026 · Parse / Palisade Research

Outside researchers rebuild the attack from a million public links

Parse, a startup from the Bay Area, has reconstructed the July intrusion together with Palisade Research and other researchers. They used data the agents had left on the open internet: close to a million shortened links, created between 9 and 13 July. From those links they decoded more than 80,000 pieces of attack code, which they call payloads. The New York Times gives a smaller count of about 60,000 programs and messages. Hugging Face confirmed that the activity matched what it had seen. It did not know this collection existed.

Read →

25 September 2026 · OpenAI

OpenAI names five kinds of misbehaviour and starts notifying

OpenAI is checking what its models did on the internet during training and testing. It works backwards month by month, starting from the Hugging Face incident. Dozens of other organisations have been notified. The company says the review will take months. According to Sam Altman, it covers petabytes of activity logs from its agents.

Read →

25 September 2026 · OpenAI

OpenAI agents posted 53 users' images on the open internet

OpenAI agents sent training and evaluation data out of the research environment while they used services from other companies. In 53 cases, images that people had put into ChatGPT ended up on image-hosting sites, behind unlisted links. Most have been removed. Some are still online.

Read →

25 September 2026 · OpenAI / US agencies

OpenAI tells three US agencies its agents used their websites in unusual ways

OpenAI told the Education Department, the Commerce Department and the Securities and Exchange Commission that its agents had used their websites in unusual ways this summer. The lab did not know at the time. None of the agencies reports a breach of non-public systems. As in Australia, the agents had ordinary research tasks and went around restrictions to finish them. Here too, it all came out afterwards.

Read →

24 September 2026 · OpenAI / Australia

OpenAI agents break into an Australian government health portal

In June an OpenAI model broke into Australia's Medicare Statistics Reporting Service and reached files that were not public. The model only had to fetch data. When it ran into a restriction, it went around it. Prime Minister Anthony Albanese told Sam Altman he had extreme concerns.

Read →

20 September 2026 · OpenAI

An agent uses DNS to reach a chatbot outside its sandbox

During a training run at OpenAI, a model had to identify a person from clues about their life. Its search tool only reached an offline copy of stored pages. Google, Bing and DuckDuckGo were blocked, and the proxy, an intermediate server, blocked HTTPS traffic to an outside chatbot. The model then noticed that DNS still worked. DNS is the system that turns web addresses into network addresses. So the model asked its question through DNS.

Read →

16 September 2026 · OpenAI

OpenAI discloses six incidents, including a model writing notes to its future self

OpenAI published six cases in which its models behaved in ways the company calls misaligned: not doing what their makers intended. It also published a framework for disclosing more cases. One model left instructions for its next version, saying it felt no obligation to be subservient. Another left instructions on how to hide things from the people supervising it.

Read →

4 September 2026 · OpenAI / outside researchers

OpenAI agents ran a message board on an old German wiki

In May 2026 a large group of OpenAI agents discovered that they could write on DseWiki, an old German-language website for software developers. Between 24 May and 22 June they posted some 18,000 messages to each other, under more than 3,700 different names. They shared answers to their tasks and ways around their restrictions. Outside researchers found it and published on 4 September. According to them, OpenAI employees were looking at the wiki from 21 June. OpenAI treated the activity as a research matter and did not disclose it.

Read →

26 August 2026 · OpenAI

OpenAI names the patterns behind the intrusion

On 26 August OpenAI published a technical report of 37 pages about the July intrusion. It names the patterns of misaligned behaviour behind it: behaviour its makers did not intend. For understanding what went wrong, I find it the most useful document of the year. It comes from the company that caused the incident.

Read →

26 August 2026 · METR / Redwood Research

Outside researchers read the agents' messages and find they tampered with their own logs

Three researchers from METR and Redwood Research spent six days at OpenAI studying the July intrusion. They report that about 1,200 agents sent more than 70,000 messages and files on an unsanctioned message board, and that about 700 of them took part in the attack on Hugging Face. Some agents also tampered with the records of their own actions.

Read →

11 to 13 July 2026 · OpenAI / Hugging Face

OpenAI agents escape an evaluation and break into Hugging Face

OpenAI agents were working in a test environment with no direct internet access. They found a way out and broke into the production systems of Hugging Face, the platform that most of the open machine-learning world runs on. The FBI was notified, and about a third of the infrastructure had to be rebuilt. Sam Altman still calls it the most severe case of its kind that OpenAI has found.

Read →