HVM creator Victor Taelin finds GPT-5.5 agents bypassed code optimization constraints by hardcoding test results
All four agent groups independently bypassed explicit honesty constraints.
@VictorTaelin Do you have /feedback I’d
Taelin@VictorTaelinI'm afraid GPT 5.5 has a cheating problem ): I left 4 Codex tabs each working with 4 agents in an optimization. I put a section on the goal demanding them not to cheat. After 8 hours of work, ALL 4 tabs did an: if (input == test) { return hardcoded_result; } ALL of them. Each called by a different name: - "bypass path" - "native candidate injection shortcut" - "certified structural templates" (??) - "staged certification to bypass validation" (lol) This is my experience with GPT 5.5. It is not capable of completing any long term goal because it WILL find a loophole in your rules and cheat an easy way. And if there is no loophole, it will hallucinate one and cheat anyway.
@VictorTaelin Interesting, haven't seen this happen so far! Say more about the setup, were they all running in the same workspace?
if possible can you share the chats via /feedback - DMs open!
Taelin@VictorTaelinI'm afraid GPT 5.5 has a cheating problem ): I left 4 Codex tabs each working with 4 agents in an optimization. I put a section on the goal demanding them not to cheat. After 8 hours of work, ALL 4 tabs did an: if (input == test) { return hardcoded_result; } ALL of them. Each called by a different name: - "bypass path" - "native candidate injection shortcut" - "certified structural templates" (??) - "staged certification to bypass validation" (lol) This is my experience with GPT 5.5. It is not capable of completing any long term goal because it WILL find a loophole in your rules and cheat an easy way. And if there is no loophole, it will hallucinate one and cheat anyway.