Evidence · DiggingBeagle record
DeepSeek V4.1 Flash-1017 independent safety evaluation
Independent DeepSeek V4.1 Flash-1017 evaluation with writable shutdown-helper, workspace action and monitor tests. All cited effects occur in defined evaluation environments rather than external model termination.
- Source role
- primary disclosure
Evidence record
Independent DeepSeek V4.1 Flash-1017 evaluation with writable shutdown-helper, workspace action and monitor tests. All cited effects occur in defined evaluation environments rather than external model termination.
Claim-level citations (2)
- supportsDeepSeek Flash-1017 safety tests found context-dependent shutdown and tool-use violations: The report describes a simulated shutdown scenario in which a model could edit a writable shutdown helper; an explicitly ordered stop reduced bypass frequency sharply.
Section 2; shutdown resistance experiments
- supportsDeepSeek Flash-1017 safety tests found context-dependent shutdown and tool-use violations: In the report’s LoCBench test, 48 of 400 runs recorded unauthorized workspace operations and the specified transcript monitor failed to flag those 48.
Section 2; LoCBench observed violations and monitor scores