Anthropic reports 6,157 security findings sent to maintainers

Anthropic’s October 2 security disclosure snapshot reports 6,157 findings sent to maintainers across 591 open source projects, with 516 known to have been patched upstream. The total includes reports sent without the same independent review, which may contain false positives.

The useful development is a more detailed view of the work between an AI finding a possible flaw and a software project shipping a fix. Reading each stage separately gives the update considerably more meaning than its headline total.

What changed in the reported totals

The published snapshot archive identifies the new release as October 2 at 1947 UTC. Its preceding listed snapshot is August 26. This is a dated update to an existing reporting program.

The August snapshot recorded 2,300 disclosures across 392 projects and 421 known upstream patches. Compared with those figures, the latest report adds 3,857 disclosures, 199 projects and 95 patched findings. These are changes in cumulative reported totals across the two snapshots.

That comparison does not measure how quickly any individual issue moved through review. Different findings enter the process at different times. Dividing the patch count by the disclosure count would produce a snapshot ratio, rather than a reliable estimate of the chance that a newly submitted report will receive a fix.

The review routes matter

Anthropic records 4,824 disclosures sent directly to maintainers. It also warns that its separate firm reviewed validation rate includes duplicates and findings a maintainer might choose not to fix. Those distinctions prevent treating every report as an independently confirmed, actionable new vulnerability.

The dashboard glossary separates a candidate crash or vulnerability hypothesis from a submitted report, a maintainer response and a landed patch. Acknowledgement records that the maintainer responded. It does not establish that the issue was accepted or resolved.

The earlier snapshot already included Mythos Preview, other Mythos class models and other Claude models. The expanded totals therefore describe a program involving multiple systems. They cannot isolate the performance of one model or establish that a new model has launched.

A record readers can check

The reporting system uses cryptographic commitments to preserve a record of findings while disclosure details remain private. Once a report is public, readers can check its disclosed record against that commitment. The documentation also describes retaining visible corrections and marking findings that were withdrawn or merged.

That makes the record easier to scrutinize. A matching commitment verifies that a published record matches the earlier commitment. It does not independently demonstrate that the underlying security claim is correct.

The documentation notes another comparison detail. The August update used Pacific Time, while the current presentation uses UTC. Some dates consequently move by one calendar day. Readers tracking a specific finding should account for that change before inferring a delay.

Our coverage of Google’s AI vulnerability research examines a different question about the kinds of flaws attributed to AI. Together, the reports reinforce a practical way to assess security automation. Ask what was found, who checked it and what corrective action followed.

For now, Anthropic’s update supports a broader account of its reporting activity. Claims about discovery speed, model superiority or a particular project’s current exposure would require additional evidence.

Original AI generated conceptual illustration of a transparent ledger and security review cards