CorrectGuard: Eyes-Off Correctness Estimation for Black-Box Security Guardrails
CorrectGuard estimates black-box AI guardrail correctness without inspecting user input or guardrail internals, boosting error-detection accuracy up to 25 points.
CorrectGuard is an eyes-off correctness estimation framework in which an independent model-based evaluator predicts whether black-box guardrail decisions are correct using only labeled eyes-on data, supporting both human eyes-off and machine eyes-off (privacy-preserving input fingerprinting) settings. It was evaluated across 13 safety and security datasets covering harmful content, jailbreaks, prompt injection, and extraction, treating open-weight guardrails uniformly as black boxes. In-context-learning classifiers improved error identification by up to 25 percentage points in macro accuracy, and finetuning-based approaches added nearly 15 points. The best correctness rankings reduced AURC from 0.33-0.44 to 0.17-0.22 across 3 guardrails while retaining 37.5-52.0% of decisions at 15% observed risk.
- Audits guardrails without access to user inputs or guardrail internals
- Evaluated on 13 datasets covering jailbreaks, prompt injection, and extraction
- ICL classifiers gain up to 25 macro-accuracy points; finetuning nearly 15
- Best rankings cut AURC from 0.33-0.44 to 0.17-0.22 across 3 guardrails
- Organizations
- CorrectGuard
From the original article
1 min read at arxiv.orgAI services increasingly rely on black-box security guardrails, yet privacy-preserving model auditing regimes often cannot measure how well these systems perform in both a human eyes-off production setting, which disallows human inspection of user input, and a machine eyes-off setting, which disallows model inspection of such input.
By Adam Faulkner · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.