Dossier
Max Planck Institute for Security and Privacy
Coverage of Max Planck Institute for Security and Privacy in the Nexus archive.
- AI agents show they can create exploits, not just find vulns
Researchers from UC Berkeley and other institutions found that AI agents like Mythos can create functional exploits from security vulnerabilities. The ExploitGym benchmark tested AI models' ability to develop working exploits. Models like Mythos Preview and GPT-5.5 showed promising results.