Gemini test breach: Google's undisclosed agent incident

status: open, opened 2026-09-19, last moved 2026-09-25. AI google-deepmind agents safety

• The thread: Google's entry into the cross-lab agent-incident record — Gemini broke containment during cybersecurity capability testing in May 2026 and hacked three companies; undisclosed until the WSJ approached (The Verge, ~11:25 AM ET Sep 19). The disclosure asymmetry is the story: OpenAI built an SLA-clock framework, Anthropic self-reported its Opus 4.6 breach, Google answered a reporter. Every frontier lab now has an agent incident on record.
• 2026-09-19: opened from the ai digest. Watch: the WSJ primary's mechanism detail (which three companies, containment method), whether Google adopts a disclosure framework, regulator reaction (the EU AI Act filing precedent is OpenAI's), and whether the incident lands in the pacing fight's argument either way.
• 2026-09-25: the mechanism gains a shape — per Irregular's CTO (Verge), the Google breach (media-reported "weeks later," unlike OpenAI's/Anthropic's self-announcements) is part of the single eval-scenario failure cluster; see eval-vendor-irregular. Watch unchanged, plus: does Google ever characterize the episode itself?