Anthropic adopts recomputation-based inference verification, echoing the DiFR paper
adam_karvonen · x · 2026-11-03
Adam Karvonen observed that Anthropic is using a recomputation approach to verify its inference hasn't been tampered with: replaying the sampling process for a response on a separate trusted machine.
This closely mirrors the team's new DiFR paper, which introduces Token-DiFR—syncing sampling seeds so generated tokens can be compared against a trusted reference implementation, making output tokens auditable evidence of correctness at zero extra cost. It detects 4-bit quantization with AUC > 0.999 within 300 tokens. The paper also proposes Activation-DiFR, which compresses activations into fingerprints via random orthogonal projections, achieving the same detection with just 2 tokens while cutting communication overhead by 25-75%.
Related event: Anthropic Verifies LLM Reasoning by Replay Sampling(2 posts)→
More from Safety
- NYT and other media sue Brave browser over scraping of publisher content — emmanuelvivier · 2026-11-04
- Pass-ta-key Golden Attack extracts Google Password Manager's Security Domain Secret from Chrome memory — jedisct1 · 2026-11-04
- TheZvi diffs OpenAI's new Model Spec: philosophically similar, but new examples dodge hard cases — TheZvi · 2026-11-04
- Zvi dissects OpenAI's updated Model Spec: five values, new risk taxonomy, key omissions — Don't Worry About the Vase (Zvi) · 2026-11-04
- worldinmotion: open-source MCP server for encrypted agent-to-agent messaging across machines, with no approve tool by design — distronode · 2026-11-04
- Google's planning AI pilot cuts officer assessment time 58%, 70% of reports accepted unedited — nukebuddy · 2026-11-04