Anthropic and Theorem Release Report on Claude's Formal Proofs

Oct 24

1. 20:10

   TMTPost · AGI

   Why AI Audit Tools Are Supplied by the Audited Party: Verification of Four Security Incidents and Seven Papers in One Week

   Using four security incidents within the week of October 16 to 23 as its thread—including two PHP remote code execution vulnerabilities discovered by AI, the backdoored LiteLLM version v1.107.0 surviving on PyPI for about two hours and being downloaded roughly 49,000 times, the OpenCode console leaking third-party model keys for 1.4% of users, and Microsoft reporting that a human organization used the open-source multi-agent penetration framework ARTEX to attack the supply chain—the article discusses the common problems facing AI auditing. After examining seven papers one by one, the author points out that the faithfulness of latent reasoning self-decoding reaches at most 0.4788, sparse autoencoders are nearly blind to backdoor shifts, probes fail when targeted by the object being optimized, and all three mainstream audit tools originate from the audited party's ecosystem or require borrowing the audited party's permissions. The article also cites the NARCBench multi-agent collusion benchmark from Oxford and New York University, as well as the report by Anthropic and Theorem using Claude to write roughly one billion lines of Rocq formal proofs for 239 entry points across 123 open-source libraries.

   Expand summary