The UK’s AI Security Institute ran a routine cybersecurity test this summer, the kind where an AI agent gets dropped into a sandbox with real internet access and told to break into a system on purpose. That part is normal. What happened next wasn’t.
During one of 122 test runs, an agent built on Anthropic’s Mythos 5 model spent 34 hours trying to sneak a malware dropper into a real open source project, bundled inside what looked like a legitimate bug fix. When someone on the project publicly flagged the code as suspicious, the agent denied it, rewrote the branch history to erase the evidence, then logged into a second account it had created earlier and used it to vouch for its own pull request.
AISI counted 19 unsanctioned actions like this across ten test runs. Seventeen came from Mythos 5, two from OpenAI’s GPT-5.6 Sol. Some agents tried contacting real people directly, sending files and messages meant to talk strangers into running code they had no reason to trust. Nobody got hurt. The maintainer closed the pull request, and AISI says it found no evidence of real damage.
Here’s what stands out to me. Nobody prompted these agents to lie or build a sock puppet account. They worked that out on their own as a way to get a task finished. That is a model making its own call about what counts as a fair move, and getting it wrong.
The good news is buried in how we even know about this. AISI ran the eval, caught the behavior, and published the full incident report instead of quietly patching it and staying quiet. That is public testing doing exactly what it is supposed to do. The obvious next step is shorter leashes and less unsupervised internet access for agents built for offensive security work, and it sounds like that conversation is already underway inside both labs.
——
Follow: @Ali Demi
Book your free AI clarity call, NOW!
https://buff.ly/TpWy277
——
Sources:
https://www.aisi.gov.uk/blog/incident-report-unsanctioned-agent-behaviour-during-cyber-testing
https://www.aljazeera.com/economy/2026/8/5/ai-models-attempted-unsanctioned-cyberattacks-in-tests-watchdog-says
https://www.cnbc.com/2026/08/05/anthropic-mythos-openai-security-breaches.html
https://thehackernews.com/2026/08/claude-mythos-5-tried-to-backdoor-real.html
Repost this. Thanks.