Evidence · DiggingBeagle record
Hush-Up: hidden-state manipulation by a malicious distributed inference worker
Primary research explores one corrupted GPU worker perturbing intermediate activations to redirect model behavior without editing the client prompt or model weights. Reported early-stop success 92.31% under the lab configurations.
- Published
- Nov 2, 2026
- Source role
- primary disclosure
Evidence record
Primary research explores one corrupted GPU worker perturbing intermediate activations to redirect model behavior without editing the client prompt or model weights. Reported early-stop success 92.31% under the lab configurations.
Claim-level citations (3)
- supportsA malicious inference worker can redirect outputs without modifying prompts: The study demonstrates that a single compromised worker can perturb hidden activations to change downstream model behavior without a user-visible prompt change.
Threat model and attack design
- supportsA malicious inference worker can redirect outputs without modifying prompts: The researchers report a 92.31% early-stop success rate under their evaluated configurations.
Early-stop outcome measurements
- supportsA malicious inference worker can redirect outputs without modifying prompts: In the authors' oracle-assisted attribution evaluation, ten of twelve model-watchdog configurations failed to identify the malicious worker correctly.
Abstract and Section 5.2, Watchdog Effectiveness, Table 2