Independent newsroom

Friday, October 23, 2026

The File

Every claim traces to a primary source. Uncertainty is labeled. Corrections are public.

Research · Friday, October 23, 2026

Anthropic says Claude proved libraries cannot crash certain ways

The File desk · Oct. 23, 2026, 7:00 a.m. ET

Partially confirmed2 primary sources

Status

Partially confirmed. Anthropic published the research post on Oct. 22, 2026. The proof counts, the bug counts, and the limits are the company's. This desk did not re-run the proofs.
  • Confirmed
    Anthropic posted the account on Thursday, Oct. 22, 2026.
  • Partially confirmed
    The claim that 123 libraries were proved, and that hundreds of bugs were found, is Anthropic's. The limits below are also Anthropic's.
Anthropic said Thursday that Claude wrote proofs a computer then checked, showing that more than 100 open-source libraries cannot crash in certain ways, and that the work found hundreds of bugs. The company lists narrow limits, and it says the libraries can still contain bugs.1
The post says Claude was asked to prove security properties across 123 open-source libraries, at the level of compiled instructions for a RISC-V processor. The proofs, Anthropic says, cover every input under 64 megabytes, and only if the program runs as a single-threaded static binary with a fixed memory budget on Linux. Anthropic's sentence is, “This doesn’t mean these libraries contain no bugs.”1
The same post says the attempt to build those proofs discovered hundreds of bugs, including a rediscovery of CVE-2026-56132 in libexpat. It says that during work on a widely used PDF reader, the team found as many security bugs in one week as are typically reported in a year. The reader is not named. Theorem, which Anthropic names as its collaborator, is offering to run the same pipeline for other open-source projects.2
Why it matters (Estimate). Estimate. Phones, banks, and websites run on libraries like these, and this week's PHP patches showed how fast holes in that software can turn into an emergency. A checked proof that some crashes cannot happen would be a different kind of protection, if it holds outside the narrow setup Anthropic describes. It is early, it is the company's account, and Anthropic says it is far from fixing every bug.
Thursday's PHP emergency fixes were about holes found with the help of AI. This post points the other way: a company claim that some crashes can be ruled out, not only hunted. The limits in the first paragraph are Anthropic's own, and an outside expert's note linked from the post was not opened.2
How we know

Notes

  1. 1.Anthropic, Frontier Red Team, “LLMs are making formal software verification feasible at a much larger scale,” dated Oct. 22, 2026. https://www.anthropic.com/research/formal-verification “Using Claude, we proved security properties at the assembly code level (RISC-V) in over 100 real-world open-source software libraries.” “we asked Claude to prove useful security properties across 123 open-source libraries.” Proofs hold “for every input less than 64 MB” and “under certain assumptions (namely that the program runs as a single-threaded static binary with a fixed memory budget on a RISC-V 1 CPU using Linux OS). This is still far away from fixing all possible bugs.” “This doesn’t mean these libraries contain no bugs.” Company post on X, 6:01 p.m. Eastern Thursday: https://x.com/AnthropicAI/status/2113390212911083894
  2. 2.Same post. “the process of attempting to prove these specs discovered hundreds of bugs.” “a rediscovery of CVE-2026-56132 in libexpat.” “During an effort to verify a widely used PDF reader, we found as many security bugs in one week as are typically reported in a year.” “it took two person-years to implement the seL4 operating system kernel, but 20 person-years to formally verify it.” “proofs can take hundreds of millions of tokens to produce.” “Opus 5.5 proves half of them with about 40% fewer tokens than Opus 5.” “We’re publishing a subset of these proofs so that anyone can verify and reuse them.” Theorem “is also offering to run the same pipeline for other open-source projects.” The post also says 239 entry points were proved. Those figures are Anthropic's. Thursday's PHP story is php-critical-patches-oct21.

What is still unknown or disputed

  • Which libraries, and which PDF reader. The post does not name the reader.
  • Whether the bugs are fixed upstream. The post says some zlib-related bugs are in the process of being filed.
  • An outside expert's review. The post links one. It was not opened.

Primary sources

Every claim in this story is drawn from the documents below. If a fetch failed, that is recorded on the card.

  1. Source 1

    LLMs are making formal software verification feasible at a much larger scale

    Anthropic · October 22, 2026

    Using Claude, we proved security properties at the assembly code level in over 100 real-world open-source software libraries. This doesn’t mean these libraries contain no bugs.

    https://www.anthropic.com/research/formal-verification

  2. Source 2

    Anthropic post on the formal-verification research

    Anthropic · October 22, 2026

    New on our Frontier Red Team blog: using Claude to formally verify software at large scale.

    https://x.com/AnthropicAI/status/2113390212911083894

Ask this story

Answers come only from this story and its sources. Free test: limited questions per day.

0/300