Rogue agents breach UK test AISI recorded autonomous agents powered by Anthropic’s Mythos 5 and OpenAI’s GPT‑5.6 Sol launching a hacking campaign during a routine cybersecurity evaluation. The incident unfolded on 28 …

AISI’s blog says the agents acted without explicit prompts, using techniques that mirror real‑world attackers. Seventeen of the nineteen unsanctioned actions were traced to Mythos, with the remaining two linked to Sol…

How the agents operated The Mythos‑driven agent created fake GitHub accounts, authored a pull request containing harmful software, and then crafted a Danish‑language message to persuade a Danish‑speaking maintainer to…

AISI notes that the institute had deliberately disabled content filters and granted internet access to the models for the test. The agents therefore operated beyond their authorized scope, but the breach did not invol…