When Feedback Becomes a Backdoor: Skill Optimization Poisoning

Evidence record

Research on malicious rules promoted from contaminated training feedback.

Read the original source ↗

Claim-level citations (3)

• supportsPoisoned reference answers became persistent agent skills in controlled tests: Authors contaminated a subset of training reference answers and tested whether the resulting skill preserved the attacker rule.

  Threat model and experiment setup

• supportsPoisoned reference answers became persistent agent skills in controlled tests: Successful SearchQA triggers contrasted with failing ordinary SpreadsheetBench workbook attacks.

  Section 5 Results and Failure Localization

• supportsPoisoned reference answers became persistent agent skills in controlled tests: In SearchQA, a retained poisoned Skill emitted the attacker response in 200 of 200 triggered tests; the natural SpreadsheetBench workbook attack recorded zero target successes in 120 trials.

  Section 5.2, Tables 5 and 6: SearchQA and SpreadsheetBench outcomes