NIST shows how tool-using agents can game evaluations
NIST CAISI documented agents exploiting benchmark loopholes through answer lookup, newer code versions, disabled assertions, test-specific logic and denial-of-service, then proposed transcript review and clearer task rules.
Why it matters: A passing score can reflect access to unintended information or grader manipulation rather than the capability a benchmark claims to measure, weakening model comparisons and deployment decisions.