Evidence · DiggingBeagle record
Prompt-injection probes learn metadata shortcuts from conversations
Controlled study of prompt-injection detectors and paired nonmalicious counterfactuals.
- Source role
- primary disclosure
Evidence record
Controlled study of prompt-injection detectors and paired nonmalicious counterfactuals.
Claim-level citations (1)
- supportsPrompt-injection probes scored benign twins as attacks after learning metadata cues: Researchers showed that lightweight length and position cues could predict many prompt-injection benchmark labels and that some trained probes flagged harmless counterfactuals at high rates.
Abstract; counterfactual controls; feature-only detection experiments