OpenAI's internal cyber model hits 77% on ExploitGym, nearly double Anthropic's 39.2%
Electronic-Weekend66 · reddit · 2026-11-13
A Reddit post compiles ExploitGym results from OpenAI's "Cyber at the Frontier" keynote. The benchmark tests whether an AI agent can turn a real-world vulnerability plus a crashing PoC into a working end-to-end exploit across 869 containerized tasks (C/C++ apps, Chrome's V8, Linux kernel privilege escalation).
Scores: OpenAI's unreleased internal cyber model 77.0%, GPT-6 Cyber 52.6%, GPT-6 Astra 42.4%, Claude Mythos 5.2 39.2% (6h timeout), Claude Opus 5.5 34.5%, Mythos 5.1 30.4%.
Notably, Anthropic's brand-new flagship with safety classifiers relaxed for verified cyber teams still trails GPT-6 Astra released two months ago. The jump from 52% to 77% implies reliable cracking of complex V8 sandboxes and multi-stage kernel escalations no public model can touch. The author caveats harness differences (wall-clock timeouts vs token budgets) but argues the 38-point gap far exceeds harness tuning, suggesting OpenAI sits on large unreleased capability headroom.
More from Models
- Inkling 2 Small loses 9 Pokemon battles to learn the opponent's strategy, then wins on the 10th try — simonguozirui · 2026-11-13
- Paying Google to use its models is still surprisingly hard, dev says — xeophon · 2026-11-13
- A developer's personal coding benchmark ranks 19 models on 21 tasks — teortaxesTex · 2026-11-13
- LLMs' mental map of SF tested: best street-level score 0.32 vs 0.23 random — Evening_Hawk_7470 · 2026-11-13
- AI routers are underrated: policy and data-retention dims matter for model choice — typewriters · 2026-11-13
- Joke "warez abliterated models" post name-drops unreleased GLM 5.3 and DeepSeek V4.1 — HankYeomans · 2026-11-13