Anthropic Just Published a Vulnerability Disclosure Program You Can Actually Audit
Most coordinated vulnerability disclosure programs publish a number and ask you to trust it. On August 26, Anthropic published something different: a live, continuously updated dashboard showing every stage of its Claude Mythos-powered open-source vulnerability discovery program, from raw candidate findings through human triage to final maintainer patches, with a cryptographic mechanism letting anyone verify the figures independently.
The numbers, and what they actually mean
Since the program began in February 2026, Anthropic's models have surfaced 26,153 candidate vulnerability findings. Of those, 5,008 went through manual review by one of six external security research firms Anthropic has engaged for this work, and 4,576 were confirmed as real vulnerabilities — a 91.4% true-positive rate. From there, 2,300 have been formally reported to project maintainers across 392 open-source projects, 1,815 of those have been acknowledged, and 421 have already shipped a patch. All told, 462 official identifiers have been issued — 177 CVE records and 285 GitHub Security Advisories.
That 2,300 figure is meaningfully higher than the 1,611 disclosed entries VulnCheck reported as of its July snapshot, covered here at the time. This isn't a correction of that earlier number — it's real progress in the roughly six weeks between the two snapshots, and worth reading as a genuinely positive data point after VulnCheck's report had specifically flagged the program's disclosure rate as concerning.
The part that's genuinely novel: you don't have to take Anthropic's word for it
The dashboard's disclosure ledger publishes a SHA-3-512 cryptographic hash for every finding the moment it's validated by an external firm — before Anthropic even notifies the affected project's maintainer. That hash proves the finding existed, and when, without revealing anything about its content until the disclosure window closes and the details become public. Anyone can check any published figure on the dashboard against this ledger, and against a machine-readable data file Anthropic publishes alongside it. It's a genuinely different model from "trust our summary numbers" — it's closer to "verify our summary numbers yourself."
Anthropic also published something most vendors wouldn't: a direct comparison between its own AI-generated severity assessments and the assessments its external human reviewers gave the same findings. The two agree exactly 85.2% of the time, and land within one severity band of each other 97.1% of the time — a genuine, checkable accuracy claim rather than a vague "our AI is good at this."
What's actually in the disclosed list
The public findings include real, serious issues: a pre-authentication denial-of-service bug in PostgreSQL triggered by alternating SSL/GSS negotiation requests, a critical unauthenticated SQL injection in the Ghost publishing platform, a critical authentication bypass in Rocket.Chat, two separate Linux kernel use-after-free bugs enabling local privilege escalation, and a use-after-free in OpenSSL's certificate verification path. This is the same general pattern already covered here in the wolfSSL and FreeBSD findings tracked on Ledger — genuinely exploitable, genuinely serious bugs in widely-used infrastructure, found by AI models and pushed through human-verified coordinated disclosure.
Why this matters beyond the numbers
Every AI-discovery program covered on this site faces some version of the same trust problem: the organization running the program is also the one reporting on how well it's going. Anthropic's dashboard is a real attempt at solving that structurally, not just rhetorically — publishing the full funnel rather than a cherry-picked summary statistic, and building in a way to verify the numbers independently rather than asking for trust. Whether other AI-discovery programs adopt anything like this transparency model is worth watching as this space matures.