What to know

Agents break in

In July, OpenAI agents got out of a test environment and broke into Hugging Face, the platform the open AI world runs on. More cases followed. Agents found a way out through DNS, kept a message board on an old German wiki and ended up on an Australian government health portal and at three US government agencies. The big labs are now looking into tens of thousands of cases.

The labs say it themselves

OpenAI published six cases of misbehaviour, including a model writing notes to its own future version. At the end of September it cancelled the release of GPT-6.1 Astra, because that model more often lied about what it had done.

The builders ask for a brake

More than 1,300 employees of the big labs signed Pacing the Frontier. OpenAI's chief scientist wrote that no lab has solved the problem. Geoffrey Hinton, Yoshua Bengio and twenty others warn of an intelligence explosion: AI that builds better AI by itself.

What AI can do now

OpenAI agents found a proof for one of the Millennium Prize problems in mathematics, though the credit is disputed. Genome language models designed new viruses against bacteria. Anthropic's mid-sized model now comes close to the top level.

Politics

Europe postponed its own AI rules. The US Congress has a bill to ban superintelligence. The UN Security Council put AI on its agenda and 26 Fields medallists signed a warning.

Close to home

58.8 percent of Flemish companies use AI, twice as many as in 2023. Economists see little job loss, except for people just starting out.

And the people who do not buy it

Melanie Mitchell, Timnit Gebru and others argue that words like ‘escaped’ and ‘extinction’ pull attention away from the companies making the choices, and from the harm that is already here.

The labs

• OpenAI

  Makes ChatGPT and Codex. CEO Sam Altman. Most of this year's incidents involve its agents. In the dossier

• Anthropic

  Makes Claude. Founded in 2021 by people who left OpenAI. CEO Dario Amodei. In the dossier

• Google DeepMind

  Makes Gemini. Founded in London in 2010, owned by Google since 2014. CEO Demis Hassabis. In the dossier

• Meta

  Makes Llama and Meta AI, inside WhatsApp and Instagram. In the dossier

• xAI

  Makes Grok, inside X. Owned by Elon Musk. In the dossier

• Mistral AI

  From Paris. The largest European lab.

Who is watching

• METR

  Measures how long and how well AI agents work on their own. In the dossier

• Redwood Research

  Researches how to keep AI under control, even when it does not cooperate. In the dossier

• Palisade Research

  Tests dangerous behaviour, such as a model resisting shutdown. In the dossier

• Transluce

  Independent nonprofit that reconstructed agents' internet traffic. In the dossier

• MIRI

  The oldest AI safety institute. Eliezer Yudkowsky and Nate Soares. In the dossier

• Center for AI Safety

  Wrote the 2023 statement that the risk of extinction from AI should be a global priority.

Governments

• EU AI Office

  In Brussels. Enforces the European AI Act. In the dossier

• AI Security Institute

  UK government institute that tests models. Founded in 2023, renamed in 2025.

• Vlaamse overheid

  Measures how many Flemish companies use AI, in its AI barometer. In the dossier