WIRED - AI Agents Are Getting Better at Writing Code and Hacking It as Well

Capture

WIRED article discussing CyberGym and broader AI-driven vulnerability discovery trends.

Security Relevance

This is useful as an external news signal for industry/public interpretation of CyberGym: AI agents are improving at software vulnerability discovery, but current systems still have major limitations and raise dual-use concerns.

Capture Summary

The article reports that UC Berkeley researchers used CyberGym to test AI models and cybersecurity agents on 188 large open-source codebases. It says the agents identified 17 new bugs, including 15 previously unknown zero-days, and notes expert caution that current systems still fall short of human bug-hunting expertise. It frames AI vulnerability discovery as both defensive opportunity and offensive risk.

Collection Notes

• News source; do not use as primary technical evidence where the paper/repository is available.
• Useful for trend context, policy/industry framing, and dual-use discussion.