Anthropic claims the team's code is entirely written by Claude and has tried automatic generation of security proofs
The story so far
The Anthropic platform engineering team stated that all code for the team and its system development and maintenance work was written by Claude, with engineers shifting to system design and verification; AI agents can now autonomously fix tests and explore directly modifying the codebase. Anthropic's Frontier Red Team partnered with Theorem, allowing Claude to convert crash tests of 123 open-source libraries in OSS-Fuzz into mathematical problems, generating approximately 180 million lines of proof code within a few weeks, which was automatically verified by Rocq, and claimed to have found defects missed by years of testing. These proofs only hold under specific processor architectures, single-threaded programs, and input sizes less than 64 MB, with a single proof consuming up to billions of tokens. Anthropic did not disclose the total cost of the project. Subsequent reports mentioned the release of Claude Code v2.1.159, and stated that 26% of Anthropic's AI R&D is led by Claude with supervision; this claim came from the report and did not provide verifiable details.
AI-written from the reports · Updated 23 days ago
Developments
3 developments
- 26 Oct, 23:25 · 1 reportsAI self-improvement researchExponential View:📈 Monday data: For science the bottleneck isn’t intelligence
- 23 Oct, 19:32 · 1 reportsClaude now writes code and proves it safe automaticallyThe Decoder:Anthropic's Claude now writes code and proves it safe automatically
- 22 Oct, 19:42 · 1 reportsCompletely adopt Claude codingITHome · AI:Anthropic's platform engineering team has fully adopted Claude for coding.
Coverage timeline
Follow the coverage to see every side of the story.
- Exponential View📈 Monday data: For science the bottleneck isn’t intelligence
Claude Code v2.1.159 released, Anthropic 26% AI R&D led by Claude and supervised; Sakana AI and Gemini-powered AlphaEvolve code agents improve SWE-bench and Polyglot by 30pp/15pp, MiniMax M2.7 model agent iteratively improves by 30%.
- The DecoderFeaturedAnthropic's Claude now writes code and proves it safe automatically
Anthropic's Frontier Red Team partnered with Theorem to convert Claude's crash tests of 123 open-source libraries in OSS-Fuzz into mathematical problems, generating approximately 180 million lines of proof code within a few weeks, which were automatically verified by Rocq.
- ITHome · AIAnthropic's platform engineering team has fully adopted Claude for coding.
All code on the Anthropic platform engineering team is generated by Claude, with engineers shifting to system design and validation, and the AI agent has autonomously fixed tests and explored modifying the codebase directly.