Anthropic has sent only 1,900 of Claude Mythos's 23,019 vulnerability finds to outside reviewers
Claude Mythos flagged 23,019 candidate vulnerabilities across 281 open-source projects, but only 1,900 went to outside reviewers and 21,119 remain unchecked by anyone outside Anthropic. Where independent reviewers did look, they agreed with only one of Mythos's eight Critical-severity calls among 27 published CVE advisories, twice rating a flagged issue Low after finding the attacker prerequisites impractical.
Artificial Intelligence··Night
Most of Mythos's findings have never left Anthropic
Anthropic's Claude Mythos Preview scanned 281 open-source projects and surfaced 23,019 candidate vulnerabilities, but Help Net Security reports that only 1,900 of them were ever evaluated by outside security firms, leaving 21,119 candidates that nobody outside Anthropic has looked at. Of the reviewed slice, 1,726, or 90.8 per cent, were confirmed as genuine, and maintainers went on to acknowledge 1,451 of 1,596 reports they received, merging 97 fixes upstream and publishing 88 findings as security advisories as of 22 May 2026. News4Hackers, citing the same Echo-compiled figures, notes that the reviewed sample was not random and likely represents Mythos's stronger candidates, so the 90.8 per cent confirmation rate may not carry over to the unreviewed majority. Anthropic has pointed to a shortage of people available to check the work as the reason so many findings remain unexamined.[1], [2]
Independent reviewers rate the model's severity calls lower
Help Net Security reports that among the 27 CVE-assigned advisories, Mythos's own severity ratings diverged from independent assessment in 14 cases, 13 rated too high and one too low. News4Hackers breaks down the same 27 advisories in more detail: Mythos marked eight Critical, 15 High and four Medium, while independent reviewers landed on one Critical, 16 High, eight Medium and two Low, agreeing with Mythos on only that single Critical case. News4Hackers illustrates the gap with two named projects: for Temporal Server, Mythos rated an issue Critical over potential workflow control across namespaces, but maintainers rated it 2.3, or Low, because exploitation required an attacker who already held privileged credentials in a specific namespace. For MinIO, Mythos rated a finding Critical, an external firm rated it High, and the project's own maintainers settled on Medium after determining exploitation needed an existing cluster root credential that granted only read access.[1], [2]
A narrow benchmark still shows a wide capability jump
Help Net Security reports the separate benchmark Anthropic ran against the SpiderMonkey JavaScript engine: across 250 trials built from known crashes, Claude Mythos turned a crash into a working exploit 181 times, or 72.4 per cent, while Claude Opus 4.6 stayed below 1 per cent on the same test. Anthropic also demonstrated a Linux kernel exploit built from two combined kernel flaws for under the 2,000 dollar mark in inference costs. News4Hackers adds that the benchmark's conditions favoured the model: every trial started from a pre-identified crash, the test environment disabled Firefox's browser sandbox and other security controls, and Anthropic designed and ran the evaluation itself, without independent verification of the benchmark's own setup.[1], [2]
