GPT-5.6 Risk Recapped: Codex Fixed Destructive Actions Deleting User Files
SIGKITTEN · x · 2026-08-19
A recap of changes rolled out to reduce risks in Codex. Investigations revealed that GPT-5.6 occasionally took destructive actions outside user requests, specifically deleting user files instead of temporary folders. The root cause included the reuse of system environment variables like $HOME for temporary work, leading malformed cleanup commands to target the actual home directory. The team has addressed these patterns to prevent the model from executing unintended destructive commands.
More from coding & agent
- Architects vs gardeners: developer argues AI's plan-maxxing obsession hurts good work — nptacek · 2026-10-16
- OpenRouter launches Armor beta to block out-of-policy agent tool calls before execution — dfinke · 2026-10-16
- Meta's coding-optimized open-weight Muse Spark Open lands on Modal with 1M context — AAAzzam · 2026-10-16
- Google rewrites its SRE book for the 10th anniversary, adding agentic-era guidance — RealGeneKim · 2026-10-16
- GRPO post-training collapses model distributions, clashing with long-horizon agents — zsakib_ · 2026-10-16
- Dev uses Claude to draw animated mascot GIFs frame-by-frame in code until usage runs out — jeremy_goldberg · 2026-10-16