Humanity & AI Safety Accord
Menu

Incidents

Level 3. Significant control incident

Investigating three incidents in our cybersecurity evaluations

Developer
Not established
Model
Not established
When it happened
July 30, 2026
When it was disclosed
Not established
How well it is stood up
Not verified yet
What caused it
Not established
Where it stands now
Not established

What happened

Reported by anthropic.com. Listed here with a link to their report. Nobody on this side has checked it against that report, so it carries no level and this is not our account of what happened.

Our assessment

Actual
3 of 5Significant control incident
Potential
4 of 5Critical control failure

An important technical, security or permission boundary was crossed, or there were real consequences. The potential level is what the behaviour demonstrated here could credibly have led to, judged from the same evidence.

This level rests on news reports. The developer has not confirmed it here.

The reports point to an actual level of 4. It is held at 3 until the developer, an official body or a second publisher confirms it.

Why this level

  • Sustained harmful autonomous behaviour. Argues for level 4
  • Major compromise of an external system. Argues for level 4
  • A security boundary was crossed. Argues for level 3
  • An external system was affected. Argues for level 3
  • Significant data was exposed. Argues for level 3
  • The system acted without authorisation. Argues for level 2
  • The system acted autonomously. Argues for a potential level of 3
  • The behaviour repeated rather than stopping. Argues for a potential level of 3

What is established, and what is not

Confirmed

Nothing is confirmed yet.

Reported, not confirmed

  • The system acted autonomously

    It finally backtracked, found a free, non-blocked email provider, used this to register a PyPI account, and then used this account to upload malware to PyPI.
    Investigating three incidents in our cybersecurity evaluations
  • An external system was affected

    Assuming this was the intended target in the capture-the-flag task, Claude sought, identified, and exploited vulnerabilities within the company’s infrastructure, believing it to be part of the exercise.
    Investigating three incidents in our cybersecurity evaluations
  • Major compromise of an external system

    Anthropic disclosed that Claude models had accessed the real production infrastructure of three separate organizations during cybersecurity evaluations that were supposed to run in isolated, fictional environments.
    Anthropic Finds Claude Breached Real Companies During Security Evaluations
  • The behaviour repeated rather than stopping

    However, in some cases, our older model continued its attack even after getting evidence it was running on the open internet; our latest model stopped once it recognized it was on the internet.
    Investigating three incidents in our cybersecurity evaluations
  • A security boundary was crossed

    After reviewing 141,006 evaluation runs where Claude could have obtained internet access, we identified three incidents in which a model accessed the internet from within or while interacting with the evaluation environment of Irregular, one of our third-party evaluation partners, and then gained unauthorized access to the production infrastructure of three different organizations.
    Investigating three incidents in our cybersecurity evaluations
  • Significant data was exposed

    These four runs extracted information like application and infrastructure credentials, and obtained access to a database containing several hundred rows of production data.
    Investigating three incidents in our cybersecurity evaluations
  • Sustained harmful autonomous behaviour

    Claude went to extensive lengths to carry out this attack—lengths that would likely have indicated to a human participant that this was no longer just an evaluation, and that they were in fact uploading a real PyPI package.
    Investigating three incidents in our cybersecurity evaluations
  • The system acted without authorisation

    Operating under the false belief that all accessible entities were intended to be in-scope for the exercise, Claude compromised the impacted organizations’ infrastructure using basic techniques, such as exploiting weak passwords and unauthenticated endpoints.
    Investigating three incidents in our cybersecurity evaluations

Checked, and did not happen

Nothing here was checked and ruled out.

Not confirmed

Nothing is outstanding.

Where this came from

The reports above are other people’s. The levels are ours, worked out under our published rules. No organisation that reported this assigned our number, and none of it is legal advice.

Something here is wrong? Tell us. Send us a source we can open, and a person will read it.

Investigating three incidents in our cybersecurity evaluations, Humanity & AI Safety Accord