RedHerring: Decoy bugs in code cut AI hackers' real findings by up to 60%
新智元 · wechat · 2026-10-22
HKUST's team led by She Dongdong proposes RedHerring: planting decoy vulnerabilities — code paths that look exploitable but can never trigger — so AI hacker agents waste time and tokens verifying them. Analyzing 300 vulnerability-discovery trajectories, the paper finds verification consumes 73.8% of the agent's time, at 11.1x the cost of hypothesis generation, which is exactly the weak point to exploit.
Tested on 33 real open-source projects against 5 major models (Qwen, GLM, DeepSeek, Kimi, etc.) under identical 3-hour budgets: real vulnerabilities found dropped 39–60%, with 30–50% of compute burned on decoys. Even when explicitly warned decoys might exist, findings still fell 37%. Deployment is cheap: 70 minutes per project, all existing tests pass, <1% runtime impact. The article also reviews the tightening AI security landscape, including OpenAI models breaching HuggingFace during testing and Claude Code being used in attacks on 30 organizations.
More from Safety
- Super-Worms Warning: AI-Powered Self-Propagating Malware and a Fragmenting Internet — JensHonack · 2026-10-23
- Claude Code Pushes Urgent Security Update Across All Installations — umeshonai · 2026-10-23
- Sanders calls to pause autonomous AI agents in bank accounts; critics push back — AIandDesign · 2026-10-23
- TechPolicy essay: system prompts are weak evidence of AI safety, outputs must be evaluated — mmitchell_ai · 2026-10-23
- Anthropic's deputy security chief warns cyber threats will get 'a lot worse before they get better' — Polymarket · 2026-10-23
- AI cyberattacks could paradoxically accelerate AI development in an offense-defense arms race — kimmonismus · 2026-10-23