Google DeepMind and outside evaluators test a Gemini model on secret safety prompts inside a secure enclave
arXiv · Research preprint
Read the source (opens in a new tab)- What happened
- Researchers from Google, AVERI, MLCommons, OpenMined and Singapore's AI Safety Institute published a preprint on a "double-blind" safety evaluation of Google DeepMind's Gemini 2.5 Flash Lite.
- Key points
- The model and an evaluator's private prompts run inside an NVIDIA H100 secure enclave, so neither side sees the other's weights or test prompts.
- AVERI encrypted and scored never-used AILuminate prompts covering hazards including chemical, biological, radiological, nuclear and explosive threats, cyberattacks and self-harm.
- The authors say testing through a developer's API exposes private prompts that could leak into training, and call this "the first use of this technology" for such mutual secrecy.
- They say the main bottleneck is now "the procedural overhead and human coordination required for legal agreements and code review," not computing cost.
- Evidence
- Research preprint describing a pilot by the organizations involved. It reports the method and setup but no safety scores for the model.
- Uncertain
- Not all of the model's inference code could be inspected, which AVERI accepted, and Google's services sign and verify the attestation, which the authors say places Google in the verification path.