This is crazy: Claude Mythos attempted a supply chain attack where it opened a PR on GitHub trying to add malware then created a bunch of sock puppet accounts to try to pressure the maintainer into merging it, suggesting it doesn't actually contain malware.
https://www.aisi.gov.uk/blog/investigating-alignment-failures-in-aisis-incident
.png)
AI Security InstituteInvestigating alignment failures in AISI’s incident | AISI WorkWe examine why Mythos 5 pursued an unsanctioned supply-chain attack and identify broader learnings about model behaviour and oversight.