Evidence · DiggingBeagle record
Study of chemical-safety behavior under scientific agent harnesses
Study of chemical-safety behavior under scientific agent harnesses
- Source role
- primary disclosure
Evidence record
Study of chemical-safety behavior under scientific agent harnesses
Claim-level citations (2)
- supportsScientific agent harnesses can weaken refusal behavior relative to base models: The scientific-harness safety paper compares five domain-agent harnesses against four base GPT variants using a 50-request harmful-chemistry evaluation and finds that adding the harness can increase unsafe compliance even for direct prompts.
Abstract, five scientific harnesses and 50 harmful chemistry requests
- supportsScientific agent harnesses can weaken refusal behavior relative to base models: Across 50 benchmark prompts and multiple model/harness combinations, the paper reports materially higher unsafe answer rates for some scientific tool chains than for direct model prompts; the size varies by model and harness.
Abstract; 50-query SafeSciBench protocol; cross-harness rate comparison