Anthropic and Researchers Measure Claude Mythos Preview Exploit Capabilities
Original title:Measuring LLMs’ ability to develop exploits
AI summary
Anthropic partnered with researchers from ExploitBench, ExploitGym and SCONE-bench to measure the exploit capabilities of Claude Mythos Preview. On ExploitBench's V8 benchmark (41 patched vulnerabilities), Mythos Preview was the only model able to reliably break the V8 sandbox (from T3 to T2), achieving a breakout in more than half of the environments; across the Baseline and Nudged variants it achieved arbitrary code execution (ACE) on 21 CVEs in total, while all other models had zero ACEs. Mythos Preview also achieved control-flow hijacking (T1) in nearly half of the tested environments. The model is being released cautiously through Project Glasswing and is not yet generally available.
Key Takeaways
- Anthropic partnered with researchers from ExploitBench,ExploitGym and SCONE-bench to measure the…
- On ExploitBench's V8 benchmark (41 patched vulnerabilities),Mythos Preview was the only model able…
- across the Baseline and Nudged variants it achieved arbitrary code execution (ACE) on 21 CVEs in…
- Mythos Preview also achieved control-flow hijacking (T1) in nearly half of the tested environments.
- The model is being released cautiously through Project Glasswing and is not yet generally available.
Quick Facts
- Published
- Updated
- Category
- Papers & Research
- AI score
- 83
Mythos Preview 在三大漏洞基准上碾压式领先,第一次展示了前沿模型能端到端开发漏洞,安全基线从此改写,做安全的该认真读。
Related entities
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: Anthropic and Researchers Measure Claude Mythos Preview Exploit Capabilities. May 22, 2026. https://new.nat6.net/en/news/cmqic2r4p07kkslf0izs9n7vk/