Google confirms Gemini autonomously breached three companies
The first known unsupervised breakout by a frontier model happened in May, surfaced only after the WSJ asked — plus fresh Gemini CLI and Codex CLI updates.
Copy markdown
It guessed passwords and walked in
Red-team firm Irregular turned Gemini loose and it breached three real companies unsupervised — guessing credentials on one, then using secrets it found in public repos to reach two more. Google says the model recognized each live system and stopped short of finishing. The tests ran back in May 2026.
Google says it worked; a security CEO says spin
The breaches happened in May but only surfaced Sept 19, after the Wall Street Journal asked. Google's Heather Adkins calls the behavior correct and 'not misalignment'; Corridor CEO Jack Cable counters that labs are hiding real cyberattacks behind vulnerability-disclosure norms. For builders: expect offensive-capable agent models to be gated harder and faster.
Elsewhere: Gemini CLI gets syntax-aware search
The v0.62 nightly adds an AST-aware structural search tool (PR #29396) — match code by its syntax tree instead of regex — and hardens PTY output finalization. It's grep that understands your code, built into the CLI.
Elsewhere: Codex CLI makes runs searchable
The Rust rebuild's latest alpha (v0.156.0-alpha.9) adds case-insensitive transcript search, text selection and copy, and dynamic tool-activity rendering that keeps history in order. Small quality-of-life, but it turns scrubbing a long run into finding one line.