Evidence · DiggingBeagle record
LLM output token choices become covert channel for stolen model weights
Primary or indexed evidence for Radar finding DB-RADAR-CP2-CP3-097. Constructed lab threat, not documented production weight theft.
- Source role
- primary disclosure
Evidence record
Primary or indexed evidence for Radar finding DB-RADAR-CP2-CP3-097. Constructed lab threat, not documented production weight theft.
Claim-level citations (3)
- supportsOutput-token choices can conceal weight exfiltration in compromised model serving: The researchers model an adversary who can perturb generated token choices to encode model-weight information in otherwise plausible responses.
Abstract, threat model and exfiltration mechanism
- supportsOutput-token choices can conceal weight exfiltration in compromised model serving: The study evaluates anytime-valid statistical tests to distinguish suspicious output changes from routine sampling variation.
Abstract, anytime-valid detection analysis
- supportsOutput-token choices can conceal weight exfiltration in compromised model serving: The proposed e-process controls false-alarm probabilities over repeated monitoring only under its stated calibration-transfer assumption; the evaluated attacks are constructed experiments.
Abstract, calibration-transfer assumption and four-model attacks