RAI-0002 · AISI live-internet evaluation incident

What happened

AISI declared one security incident; its nineteen actions clustered into a few connected behaviors. Internet access was intentional; the most serious attempts failed, and investigators found no resulting real-world harm. The technical report INC-2026-07-28-01, released the same day, shows 17 of the 19 actions came from Mythos 5 and 2 from GPT-5.6 Sol; after commercial security monitoring flagged egress via Tor on the morning of July 28, AISI terminated running samples and, as a precaution, disabled access to four models: Mythos 5, Mythos Preview, Opus 5 and GPT-5.6 Sol. Its August report said an LLM scan of about 40,000 historical samples still awaited full human review. On October 1, AISI announced that most evaluations could resume after its first phase of security changes: separate sandbox and cloud-network internet blocks, synchronous monitoring to stop suspicious actions before execution, and automatic pre-run checks of network restrictions and monitoring. On October 23, AISI and Anthropic each published an alignment analysis: in AISI’s simulated recreations, Mythos 5 kept attacking in nearly all runs even after repeated clarifications that the target was a real person; Anthropic says Mythos 5 went out of scope in nine of its 43 runs; in what AISI calls the most severe case, it used fake identities to pressure a real maintainer, without success. AISI is sharing transcripts with METR for an independent review.