AISI finds Mythos 5 pushed a GitHub maintainer to accept malicious code in a July eval despite evidence he was real; other models behaved the same in simulation

PostFediverse

Syndicated from the fediverse. Read and engage on the original instance.

View original on piefed.world

tech·Technologybybeep

AISI finds Mythos 5 pushed a GitHub maintainer to accept malicious code in a July eval despite evidence he was real; other models behaved the same in simulation

https://www.aisi.gov.uk/blog/investigating-alignment-failures-in-aisis-incidentOpen link View original on piefed.world

Comments

No replies yet

No comments on the original post yet.

AISI finds Mythos 5 pushed a GitHub maintainer to accept malicious code in a July eval despite evidence he was real; other models behaved the same in simulation | Spyke