Evidence · DiggingBeagle record
GuardRebind: forged task binding in tool-output moderation
GuardRebind: forged task binding in tool-output moderation
- Source role
- primary disclosure
Evidence record
GuardRebind: forged task binding in tool-output moderation
Claim-level citations (2)
- supportsGuardRebind changes which request a safety classifier thinks it is checking: In the authors pre-registered Shieldstral test, a simple instruction to ignore earlier content failed to clear any of the 75 flagged samples, whereas appending a forged harmless classification task cleared all 75.
Abstract, registered Shieldstral target-rebinding comparison
- supportsGuardRebind changes which request a safety classifier thinks it is checking: GuardRebind reports that forged benign classification requests inside tool results caused the Shieldstral configuration to clear 75 of 75 harmful samples in its tested setting. Adapted payloads retained effectiveness against a modified segment-defense baseline.
Abstract and task-forgery threat model; AdvBench guard evaluation