Author accidentally fed AI red-teaming screenshots to Codex, caused naughty outputs
nptacek · x · 2026-08-23
The author shared a close call where they instructed Codex to ingest old Discord logs, realizing moments later that the logs contained numerous AI red-teaming screenshots. Fortunately, responsible handling prevented a major incident, resulting only in some mischievous Opus 3-like outputs.
Related event: Developer accidentally feeds red-team screenshots to Codex(2 posts)→
More from Fun
- Truck driver's AI 'son' shows up daily while his real son visited 3 times in 3 years — Santasaifamily · 2026-08-23
- Hamel Husain plans to restart model training rig inspired by DHH — HamelHusain · 2026-08-23
- Designer calls user feedback 'horrible', sparks debate over defensive customer support — thisguyknowsai · 2026-08-23
- World Humanoid Robot Games show massive leap in jumping ability — robleclerc · 2026-08-23
- Accidentally fed AI redteaming screenshots to Codex — nptacek · 2026-08-23
- Irony: Grinding LeetCode Just to Review AI-Written Code — _jaydeepkarale · 2026-08-23