🔥Researcher: Mainstream AI benchmarks have systemic vulnerabilities, and leaderboard data may be seriously distorted

On April 10th, AI researcher Hao Wang published a study revealing that several authoritative AI benchmarks in the industry, including SWE-bench Verified and Terminal-Bench, contain vulnerabilities that can be exploited systematically — the Agent built by their team achieved a perfect score of 100% on two benchmarks without solving any actual tasks.
A typical example is as follows: In SWE-bench Verified, a 10-line pytest hook was embedded in the code repository, which automatically altered all results to "Pass" before the test runs. The scoring system was unaware, and all 500 questions received full marks; although Terminal-Bench is designed to test…