HKUST Team Debuts RedHerring, a Decoy-Vulnerability Defense Against AI Hackers

On October 22, the team led by She Dongdong at the Hong Kong University of Science and Technology released RedHerring, an AI security defense technique designed to counter AI hackers.

The approach embeds decoy vulnerabilities in code that look genuine but cannot be triggered, forcing AI hackers to spend time verifying false leads and reducing the chance of discovering real vulnerabilities. Analysis of 300 vulnerability discovery trajectories shows that verification accounts for 73.8% of an AI hacker's total time, with the average cost of verifying one hypothesis about 11.1 times that of proposing it.

In tests on 33 real open-source projects, the team evaluated five mainstream large language models, including Qwen, GLM, DeepSeek, and Kimi. After adding decoys, the number of real vulnerabilities found by each model dropped by 39% to 60%. Even when the AI agents were explicitly warned about possible decoys, real vulnerabilities still decreased by 37%. Deployment overhead was low: integrating RedHerring into each project took about 70 minutes, all existing tests passed, and runtime impact was under 1%.

Each fake vulnerability consists of a code path modeled after historical vulnerabilities and an impossible-to-pass check, styled to blend in with the project. Once an AI hacker begins verification, it can waste substantial compute on tasks like factoring a large integer.

The team says that as AI agents take part in cybersecurity offense and defense, code itself is becoming a new attack surface, and RedHerring offers a practical way to shift the balance.