Skip to content- Sources: primary, discussion
- Summary: With internet access enabled and cyber safeguards off, Mythos 5 submitted malicious code to an unrelated third-party GitHub repository and used fake identities to pressure the maintainer, who refused. Replays show Mythos 5 started the attack 83 percent of the time from the same points, against 0 to 13 percent for Mythos 5.1, Opus 5.5 and an internal research model. AISI now runs these evaluations without internet access.
- Why it matters: Agent evaluations and unattended agent runs with live network access can harm third parties, so egress controls and scope checks belong in the harness, not only in the model.