Anthropic's OSS Scanner routes Mythos findings to maintainers ahead of its own review queue
Anthropic's OSS Scanner sends maintainers Mythos findings before its researchers review them, with about 6,000 of 29,000-plus candidates triaged. The service is free and opt-in, and reports carry reproduction steps and proposed fixes where available.
The Engineer · Build desk
Key figures: nearly 5,000 reports sent in response to maintainers' requests for the remaining findings; more than 29,000 candidate vulnerabilities found over six months; approximately 6,000 manually reviewed and triaged.
| Figure | Value | Note | Claim |
|---|---|---|---|
| Reports sent on request for remaining findings | 5,000 reports, nearly | 5 | |
| Candidates found, over | 29,000 candidates | 3 | |
| Manually triaged, about | 6,000 candidates | 3 |
What happened
- Some maintainers who received Anthropic's initial reports asked it to send the remaining findings for their projects without waiting for validation.
- In a test of an early version, penetration testers examined 97 high- and critical-severity findings across 48 projects: 85 met the disclosure criteria, 11 more were real duplicates, and one was invalid.
- wolfSSL's Todd Ouska reported that 72 of 74 findings were valid, with five becoming CVEs.
- In April, the Firefox team shipped fixes for 271 vulnerabilities found during an early Mythos Preview evaluation.
Why it matters
- What it costs cost Checking severity, reachability and threat-model fit moves to the maintainers who opt in, because Anthropic's researchers no longer review each report first.
- What doesn't add up contradiction Anthropic's dashboard counts duplicates and fixes maintainers may decline as true positives, and calls patch counts the more reliable measure, so a high hit rate can run ahead of what a project ships.
- What it limits constraint Absorbing a full scan can displace other work. Mozilla's Bobby Holley warned that other projects might need to reprioritize everything else.
- What it sets up precedent Anthropic names automated triage and patch development as its long-term direction, so maintainers should expect the handoff to move further upstream of human review.
Anthropic's own numbers show how much is left. It says it found more than 29,000 candidate vulnerabilities in six months and manually triaged about 6,000 [3]. That leaves at least 23,000 unreviewed [19]. Nearly 5,000 reports have already gone to maintainers who asked for the rest [5], so roughly 18,000 findings remain, assuming every report sent came from the unreviewed pile [22]. The inputs are approximate, and so is the result.
The scanner changes who sees that remainder. Participation is opt-in [1]. According to Socket's reading of the extended FAQ, the service is aimed at projects already keeping pace with verified high- and critical-severity reports, with spare capacity to look into more [18]. A maintainer controls the volume by choosing whether to sign up.
The quality figure needs the same care. The early-version test works out to 88 percent of findings meeting the disclosure criteria [20], and 99 percent real once duplicates count [21]. For that to describe a given inbox, the unreviewed stream would need the same severity mix and project mix as the 97-finding sample. Socket notes that the evaluation gives no false-positive rate that holds across all severity levels, or across every project that signs up [7]. Anthropic also concedes that maintainers have pushed back on overstated severity ratings and on findings that misread their threat models [8].
Diminishing returns cut the other way for projects that have already run other tools. Before curl received its first Mythos report, several AI tools had prompted roughly 200 to 300 bug fixes over eight to ten months, curl maintainer Daniel Stenberg reported in May [13]. He cautioned against judging tools by bug count, because the first scans had easier pickings and the later ones were working through code that earlier rounds had already improved [14]. We'd expect the heaviest load on a project's first scan and a lighter one on the scans after it.
Anthropic frames the service as a fix for remediation, not discovery. Its Cyber Mission announcement says of Project Glasswing: "we haven't yet achieved a sufficient reduction in cyber risk" [16]. The company says it often saw months pass between discovery and a fix [17].
What to watch
- Whether Anthropic publishes false-positive rates by severity level or per-project report counts as more projects join.
- Whether maintainers outside the early group report duplicates and severity disputes at a higher rate than the 97-finding test showed.
- Whether Anthropic adds automated triage or patch drafting to the service, as its long-term plan describes.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence55
- Adoption30
- Hype gap+15
- Incentives55
- Confidence58
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Anthropic launched OSS Scanner, a free, opt-in service that uses its strongest Claude models, including Mythos, to look for vulnerabilities in open source projects.
- [2]
Participating maintainers receive reports showing how to reproduce the vulnerabilities, along with proposed fixes when available, without waiting for Anthropic's researchers to review the findings.
- [3]
Over six months, Anthropic says it found more than 29,000 candidate vulnerabilities but manually reviewed and triaged approximately 6,000.
- [4]
Some maintainers who received initial reports asked Anthropic to send the remaining findings for their projects without waiting for validation.
- [5]
Anthropic has already sent nearly 5,000 reports in response to maintainers' requests for the remaining findings.
- [6]
Anthropic tested an early version by having penetration testers examine 97 high- and critical-severity findings across 48 projects. Of those, 85 met its coordinated disclosure criteria, eleven more were real issues that duplicated known problems or other findings from the scan, and one was invalid.
- [7]
The evaluation does not establish a false-positive rate for every severity level or every project that joins the service.
- [8]
Anthropic acknowledges that maintainers have challenged inflated severity ratings and findings based on a misunderstanding of their threat models.
- [9]
Anthropic's vulnerability disclosure dashboard has a true-positive metric that includes duplicates and issues maintainers may decline to fix, including findings outside a project's threat model or in code that is not typically reachable. The dashboard describes patch counts as a more reliable, though delayed, measure of impact.
- [10]
wolfSSL's Todd Ouska reported that 72 of 74 findings were valid, with five becoming CVEs.
- [11]
In April, the Firefox team shipped fixes for 271 vulnerabilities identified during an early Mythos Preview evaluation.
- [12]
Mozilla's Bobby Holley described a team working around the clock and warned that other projects might need to reprioritize everything else to tackle the findings.
- [13]
In May, curl maintainer Daniel Stenberg reported that several AI tools, including AISLE, Zeropath and OpenAI's Codex Security, had prompted roughly 200-300 bug fixes over the preceding eight to ten months, before curl received its first Mythos report.
- [14]
Stenberg cautioned against judging tools by the number of bugs they uncover: earlier scans had more easy targets, while later ones examined code that had already been improved.
- [15]
OSS Scanner is the first step in a larger long-term effort to automate vulnerability triage, develop patches, and research ways to harden or rewrite code.
- [16]
In its Cyber Mission announcement, Anthropic said of Project Glasswing: "we haven't yet achieved a sufficient reduction in cyber risk."
- [17]
Anthropic says it often saw months pass between discovery and a fix.
- [18]
The extended FAQ explicitly targets projects that can already keep up with verified high- and critical-severity reports and have capacity to investigate more.
- [19]
At least about 23,000 candidate findings were not manually triaged.
- [20]
85 of 97 evaluated findings is about 88 percent.
- [21]
Counting the eleven real duplicates, 96 of 97 evaluated findings were real, about 99 percent.
- [22]
Roughly 18,000 findings remain unsent, if all of the nearly 5,000 reports sent came from the untriaged pile.
Get this ledger inside ChatGPT. Ask about any of our stories and the answer brings its figure, each claim’s verdict and the sources. Free, no account.
Sources
1 independent publisher whose own reporting we read for this story.
- socket.devAnthropic Gives Open Source Maintainers a Fast Track to Mythos Vulnerability Findings
1 article · October 15, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.