Artificial Intelligence
Anthropic and Theorem proved security properties across 123 libraries
Anthropic said it and Theorem used Claude to prove security properties at the RISC-V assembly level for 123 open-source libraries, while the process also surfaced bugs in some code.
- Anthropic said it and Theorem used Claude to prove security properties for 123 open-source libraries.
- The company said the proofs covered 239 entry points and were written against RISC-V binaries under assumptions that included single-threaded static builds with fixed memory budgets.
- Anthropic said Claude found bugs during the work, including an out-of-bounds write and an out-of-bounds read.
- Anthropic said it would publish part of the proofs and supporting semantics, and that Theorem would offer the same pipeline to other open-source projects.
Anthropic said it worked with Theorem to use Claude on 123 open-source libraries and to prove security properties at the RISC-V assembly level. It said the project covered 239 entry points and was the first large-scale effort of its kind.
The company said it turned existing fuzz tests from OSS-Fuzz into specifications and checked them against models of RISC-V and Linux process behaviour. Those proofs assumed single-threaded static binaries with fixed memory budgets and inputs below 64 MB.
Anthropic said Claude patched bugs it encountered and then re-checked the updated code. The examples it gave included memory errors in a decompressor and an image decoder, plus a rediscovered vulnerability in libexpat.
Anthropic said it was publishing a subset of the proofs and the supporting semantics, and that Theorem would run the same pipeline for other open-source projects. It also said it was releasing a Claude skill for developers who want to verify their own code.
Named in this story
Companies
- Anthropic
- published the post and said it used Claude with Theorem
- Theorem
- collaborated on the verification pipeline and provided semantics
- Airbus
- was cited as an example of a mission-critical user of formal verification
Organisations
- OSS-Fuzz
- supplied fuzz tests for the projects
Products and systems
- Claude
- generated proofs and patched bugs in the libraries
- RISC-V
- was the instruction set for the compiled binaries
- Linux
- was the operating system model used in the proofs
- Rocq
- checked the formal proofs
- AddressSanitizer
- made memory-safety errors crash during fuzzing
- libexpat
- was the library in which a vulnerability was rediscovered
- seL4 operating system kernel
- was cited as an example of expensive formal verification
How the source tells it
The piece read like a promotional research announcement: urgent about software risk, confident about Claude’s capabilities, and optimistic about formal verification.
- alarm sweeping claims that bugs are everywhere and current defences are not enough
- hype self-authored capability claims and a first-of-its-kind framing for the project
- urgency imperative language presenting formal verification as immediately necessary
- optimism forward-looking certainty that the approach will become essential