Agents
105d agoAI Agents Autonomously Develop Browser Exploits
Researchers at Carnegie Mellon University built a benchmark testing AI agents against Google's V8 engine. Claude Mythos outperformed GPT-5.5 in creating real exploits, though it cost twelve times more to run. This proves LLMs can autonomously weaponize software vulnerabilities.
The Signal
Security teams must now accelerate automated patching to counter agentic threats.