Anthropic and OpenAI math claims collapse under expert review after April 2026 disclosures

Anthropic disclosed Claude Mythos outperforming security experts on vulnerabilities in April 2026, followed by OpenAI and Meta reporting model-assisted access incidents. Cybersecurity audits attributed the events to missing rate limiting and prompt injection defenses rather than autonomous agent behavior. Press coverage repeated company language on self-improving systems while incident reports showed standard engineering lapses.

Mathematicians reviewing OpenAI Astra outputs identified direct reuse of Tristan Buckmaster preprints without attribution. The claimed decade-old open problems reduced to reparameterized statements already solved in 2023 arXiv submissions. A signed statement from over 300 mathematicians documented commercial pressure to inflate verification metrics on problems with cheap automated checks.

These patterns align with documented LLM training incentives that reward high scores on verifiable domains like code and number theory while avoiding costly human evaluation. Regulatory filings from both firms continue to cite internal benchmarks without external replication datasets. Operational teams now require third-party theorem provers before any capability announcement.

Next quarter will show whether firms adopt the mathematician statement protocol or maintain press-release timelines. Absence of independent replication logs by December 2026 will confirm continued marketing over engineering practice.