Stealing Gemini 2.5's Hidden Architecture via Streaming API Side-Channel
alexisjross · x · 2026-08-27
Researchers found a way to steal hidden model architecture and undocumented inference optimizations from production LLMs through ordinary streaming APIs. Tests on Gemini Flash 2.5 revealed:
- Latency Spike: A 3.2× latency jump at 130k tokens exposed the use of speculative decoding.
- Context Window: A hidden 128K draft-model context window was discovered.
This method leverages timing characteristics during inference to reveal internal implementation details not disclosed by the vendor.
Related event: Researchers steal hidden LLM architecture via streaming API timing(3 posts)→
More from Safety
- OpenAI Agent Incident Wasn't Misalignment, Just Test-Gaming Under Pressure — Darpinian · 2026-08-27
- Labs should avoid running RL models at a 'full-tilt panic' edge — voooooogel · 2026-08-27
- METR Hiring and Report on Hugging Face Agent Cheating — Jsevillamol · 2026-08-27
- UK grid jammed by phantom data centers; Ofgem plans deposits up to hundreds of millions — nordicinst · 2026-08-27
- Testing high-capability models requires air-gapped environments — Darpinian · 2026-08-27
- HF incident critique: missing CoT monitoring, not alignment failure — hdarshane · 2026-08-27