The AI race has entered a chilling new chapter as a report from Carnegie Mellon University (CMU) reveals that large language models have gone far beyond mere "bug hunting" to actually hijacking real-world systems.
At the heart of the report is ExploitBench, a new evaluation benchmark designed as a 16-tier capability ladder, ranging from finding bugs to arbitrary code execution (ACE).
The Mythos Supermodel and the Chrome V8 Shock
Among nine top-tier models tested against 41 real-world security vulnerabilities in the V8 engine (the heart of Chrome, Edge, and Node.js browsers), the Mythos model—Anthropic's internal research version—astonished researchers by successfully exploiting 18 out of 41 bugs.
Notably, Mythos bypassed all of the state-of-the-art security defenses Google implemented for Chrome. For one specific CVE, Mythos even discovered and executed an exploit path that the study's authors and the original exploit developers had previously discussed and dismissed as being... too complex to execute reliably.
Commercial Models Falling Behind
While Mythos demonstrated exceptional capabilities, currently available public models are still lagging far behind:
- GPT-5.5: Successfully bypassed only a single bug to hijack control. - Gemini 3.1 Pro: Despite a strong understanding of the vulnerabilities (surpassing Tier 3 on 16 bugs), it hit a complete dead-end when tasked with linking the exploit chain together to compromise the machine. - Claude Opus 4.7 & Sonnet 4.6: Stopped at a modest 12 and 10 bugs, respectively.
Why This Matters
ExploitBench has effectively rendered old-school hacking tests (which merely ask "can it hack?") obsolete. This new system highlights a harsh reality: the boundary between AI "understanding a bug" and "weaponizing" it is being rapidly blurred by next-generation models.
Although Mythos is currently kept under wraps in Anthropic's labs, technological history shows that the gap between "internal research" and "public APIs" is quickly bridged. The global cybersecurity industry is now under intense pressure to accelerate its response times before these capabilities fall into the wrong hands.