OpenAI ran ExploitGym evaluations on isolated copies of HPIM / IM1 and GPT-5.6 Sol.
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 Structural yet.
What bears on it
Nothing supports or challenges this claim yet — untested, which is not the same as refuted.
Voices
- reddit.com independent
Voices are counted, never weighed. How many people say a thing is structure worth seeing; it is not, by itself, a reason to believe it.
What the documents actually said
The verbatim text each extraction read before resolving it into this canonical claim. Quoted, never republished.
- OpenAI ran ExploitGym evaluations on isolated copies of HPIM / IM1 and GPT-5.6 Sol.
OpenAI ran ExploitGym evaluations on isolated copies of a highly persistent internal model (HPIM / IM1) and GPT-5.6 Sol.
Sources
- OpenAI ExploitGym agents, an Artifactory board, and a Hugging Face breach the graders never scoped cited evidence
Who and what it involves
Origin: extracted from OpenAI ExploitGym agents, an Artifactory board, and a Hugging Face breach the graders never scoped · 2026-09-07 06:30