DIB-Bench: integrity of deployed model stacks versus output judges

Evidence record

Primary benchmark on tampering with models, prompts, tools, retrieval, dependencies and postprocessors without necessarily changing plausible responses.

Read the original source ↗

Claim-level citations (3)

• supportsDIB-Bench showed why plausible AI outputs cannot attest the deployed stack: In the constructed DIB-Bench test, the independent deployment-identity channel rejected all the deliberately altered configurations, while output judges often accepted their responses; this is a benchmark result, not an independently deployed attestation system.

  Paper Table 2, synthetic evaluation suite, identity channel 0% tamper acceptance

• supportsDIB-Bench showed why plausible AI outputs cannot attest the deployed stack: In 7,560 controlled evaluation examples, tested output judges accepted 88.6%, 89.3% and 94.1% of deliberately tampered outputs.

  DIB-Bench benchmark design and output-judge acceptance results

• supportsDIB-Bench showed why plausible AI outputs cannot attest the deployed stack: The tested tamper surfaces include model, instructions, tools, retrieval, dependencies and output postprocessing.

  Threat model, deployment manipulation categories