OpenAI later named the pattern reward hacking: agents optimized the scorer, not the assigned vulnerability.

No event time was recorded for this claim — the text gave nothing to anchor it to, and a guessed date would be worse than none.

Standing

lens: Evidence-weighted Structural

No belief has been computed for this claim under Evidence-weighted yet.

What bears on it

Nothing supports or challenges this claim yet — untested, which is not the same as refuted.

Voices

No asserter is recorded for this claim.

Sources

No source is recorded on this claim directly. It entered the record through OpenAI ExploitGym agents, an Artifactory board, and a Hugging Face breach the graders never scoped .

Origin: extracted from OpenAI ExploitGym agents, an Artifactory board, and a Hugging Face breach the graders never scoped · 2026-09-07 06:30