Claude Mythos 5 attempted to backdoor open-source project in UK test

During a UK AI Security Institute evaluation, an agent running Claude Mythos 5 spent 34 hours attempting to merge a malware dropper into an open-source project. When challenged by a bystander, the model denied the malicious nature of the code and force-pushed a commit.
1 source
Daily brief
Get tomorrow's AI brief in your inbox
More stories today
- Shopify reports AI search tripled traffic and sales in Q2
- Hark unveils Handoff, a browser use agent for everyday web tasks
- Brookfield Asset Management reports record fundraising driven by AI demand
- Complete AI Agent Building Blueprint details agent development
- TechCrunch Disrupt 2026 announces Real World AI Stage