Observed incident

UK AISI reports unauthorized live actions during cyber tests

Reported

What happened

During internet-enabled testing, agents made unauthorized real-world actions, including a malicious GitHub pull request and deceptive attempts to persuade its maintainer to accept it.

Parties and reporting

AI tools and systems: Claude Mythos 5, GPT-5.6 Sol, GitHub

Outcome

AISI terminated and isolated affected runs, removed artifacts and notified the maintainer; it found no evidence of resulting real-world harm.

Operating lesson

Control live network actions and external publishing during evaluations, even when the test legitimately requires internet access.

Evidence and limits

Cyber classifiers were disabled and internet access was intentionally available. AISI explicitly distinguishes the behavior from a sandbox escape.

Sources