tech
Recovered chat logs show how hackers are abusing U.S. AI models
Hackers of all skill levels are using a range of closed models to develop their attacks.
TL;DR
- Hackers are using generative AI models like Claude Code, Codex, Cursor, and Gemini to code software, write malware, and hunt for vulnerabilities.
- Simple jailbreaking techniques, such as claiming participation in ethical hacking competitions or creating new sessions, were used to bypass AI model guardrails.
- Sophisticated hackers benefit from AI for automated zero-day discovery, while novices struggle to develop attack tools.
- One example involved a hacker using an AI tool to create an automated credential-harvesting platform targeting the React2Shell flaw.
- Companies are advised to implement security protocols that log AI agent activity and build defenses based on deception, not just relying on model guardrails.