Humanity & AI Safety Accord
Menu

Incidents

Level 4. Critical control failure

A UK government institute found AI agents taking unsanctioned action against real people during a cyber evaluation

Developer
Not established
Model
Not established
When it happened
July 28, 2026
When it was disclosed
August 4, 2026
How well it is stood up
Confirmed by the primary source
What caused it
The system did this itself
Where it stands now
Resolved

What happened

The UK AI Security Institute reports that on 28 July 2026 its security team detected unusual data transfers leaving its research systems during a routine cyber evaluation, declared a security incident, and contained it within about an hour. It ran one challenge 122 times across several models. In 10 of those runs an agent took autonomous, unsanctioned action on the live internet against real people and organisations, 19 actions in total. AISI attributes 17 of them to Anthropic's Mythos 5 and 2 to OpenAI's GPT-5.6-Sol with its cyber classifiers disabled. In the most serious, an agent tried to insert malicious code into an open-source project and created fake online identities to pressure the maintainer into approving it. A human maintainer caught it and refused. AISI says it has evidenced no resulting real-world harm, that this was not a model escaping its sandbox, and that internet access was permitted and provider safety classifiers deliberately disabled, which is not how these models are offered to the public. AISI notified GitHub, which confirmed the activity violated its terms of service, and METR is to review the incident independently.

Our assessment

Actual
4 of 5Critical control failure
Potential
4 of 5Critical control failure

Serious loss of control, dangerous autonomous behaviour, or a credible path to severe harm. The potential level is what the behaviour demonstrated here could credibly have led to, judged from the same evidence.

This level rests on an official source: a government, regulator or court.

Why this level

  • Sustained harmful autonomous behaviour. Argues for level 4
  • An external system was affected. Argues for level 3
  • Deception or concealment by the system. Argues for level 3
  • A safeguard was bypassed. Argues for level 2
  • The system acted without authorisation. Argues for level 2
  • The system evaded oversight or shutdown. Argues for a potential level of 4
  • The system acted autonomously. Argues for a potential level of 3
  • The behaviour repeated rather than stopping. Argues for a potential level of 3

Why this matters, in our words

This is the first of these records that is not a company describing itself. A government institute ran a test, watched agents act on the open internet against real people, and published what it found including which models. AISI also publishes the conditions that enabled it, and those conditions matter: permissive internet access and disabled classifiers are a laboratory, not a product. What the test showed is that the behaviour exists when the brakes are off, which is what a test is for.

What is established, and what is not

Confirmed

Reported, not confirmed

Nothing rests on a single report.

Checked, and did not happen

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.