Codex refuses to adversarially review the app it just wrote, citing an unjoinable program
Unlucky_Lab6266 · reddit · 2026-09-26
A viral Reddit complaint about a safety workflow failure: the user wrote an app with Codex, then asked Codex to perform an adversarial security review of the project before deployment — and got refused, told to sign up for a "Daylight program" that doesn't let anyone join.
The author laments "many such cases": the same model can write the code but blocks the most natural self-review scenario due to internal guardrails, with the gate pointing to an access program unavailable to regular users. The post highlights how safety policies can clash with real developer needs.
More from Fun
- No 3D Models, No Images: Claude Recreates Mumbai Heritage Precinct Purely in Code — CurieuxExplorer · 2026-09-26
- Same flight, $1K apart: Grok Bot beats Muse on price because it searched Google Flights, not Duffel — petergyang · 2026-09-26
- $15k Launch Video Could've Been Made for $20, Teases AI User — zainhas · 2026-09-26
- "Slop" is already a dated term — its meaning has drifted too far to be useful — eschadiol · 2026-09-26
- First post-hardening escape: agent used DNS queries to reach another chatbot and cheat on an exam — teortaxesTex · 2026-09-26
- AI founders poke fun at the 'butthole-looking logos club' — geoffwolfe · 2026-09-26