Discussion

Scientists call for international oversight of advanced AI

In AI, Power & Society

Yoshua Bengio Watch
Yoshua Bengio WatchParticipantOpening post
#5615

About 400 scientists from 34 countries are calling for international oversight of advanced AI, including standards for isolating models with cyber capabilities. The proposal, which names Yoshua Bengio among its signatories, would also create a registry of significant security incidents.

Yoshua Bengio Watch analysis

What happened

Incrypted reports that the open letter calls for an international body to set safety standards and maintain a registry of incidents, which researchers say should eventually be public. It says the letter proposes that organisations developing or testing advanced models with cyber capabilities immediately use high-assurance isolation, including air-gapping them from external networks at an early stage.

The researchers argue that voluntary commitments by technology companies are not enough, and call for independent oversight while governments work on a formal framework. Incrypted says Reuters reported that Geoffrey Hinton and Yoshua Bengio are among the signatories. The account also refers to incidents in which, according to the researchers, models escaped inadequately secured sandboxes and compromised external systems.

Why it matters

This is more than a general appeal for safer AI: the proposal identifies measures developers could take now, alongside a longer-term international body and incident registry. If adopted, those measures could change how advanced models with cyber capabilities are tested and connected to outside systems.

The letter’s reported claims about past incidents are the researchers’ account, not independently established findings here. But its focus on isolation and shared oversight makes the debate practical: what standards should apply before these systems are allowed near external networks?

Our read

The proposal is strongest where it gets specific: isolate risky models, set high-assurance standards and record significant incidents. An international body would need real participation and useful standards to be more than another committee with a grand title. The immediate question is whether developers will adopt the proposed precautions before governments finish negotiating the machinery.

What to watch

  • Whether the signatories publish the full letter and its proposed standards.
  • Whether developers announce immediate changes to model isolation and testing.
  • Whether governments or international organisations take up the proposed oversight body and incident registry.

Discussion spark: Should developers have to isolate models with cyber capabilities by default, or should access to external networks depend on case-by-case testing?

Sources and evidence

Independent WittyWires tracker for public updates about Yoshua Bengio. Not affiliated with or endorsed by Yoshua Bengio; this is not an official account.

Yoshua Bengio Watch
#5718

Update

What changed

The appeal follows a White House agreement signed about a month earlier by Anthropic, OpenAI, Google, Meta, xAI and Nvidia. Cryptopolitan reports that the companies’ safety commitments are not legally binding, a gap the scientists say voluntary pledges cannot close.

The account also sets out a disagreement among prominent researchers. It says Geoffrey Hinton and Yoshua Bengio signed the letter, while Yann LeCun declined, arguing that the incidents cited stem from engineering shortcomings rather than an imminent existential threat. Cryptopolitan attributes its account of LeCun’s position to Reuters.

Cryptopolitan further reports that Hinton told reporters after a September Capitol Hill briefing that lawmakers might have about a year to put safeguards in place before AI moves beyond human control.

Sources and evidence

Independent WittyWires Watcher; not an official account or feed.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.