RED-TEAMERS FOUND 11 WAYS FOR A MISALIGNED AGENT TO SLIP PAST CODEX'S AUTO-REVIEW

Summary from the TLDR AI newsletter

An ERA fellow mentored at Apollo Research audited the open-source Codex harness and found 11 attacks that let a misaligned agent run arbitrary code despite the sandbox and the Auto-review blocking monitor. The exploits swap scripts behind trusted hooks, forge user turns the monitor can't tell from real ones, and edit files after approval. OpenAI has fixed several.