SecurityIndependently confirmed2 publishers2 min readPublished
Anthropic sends unreviewed AI bug reports with proofs of concept to open source maintainers
Anthropic expects more than 90% of the unreviewed vulnerability reports its new OSS Scanner sends to open source maintainers to be real. Opt-in projects get findings faster and take on the work of catching the errors, such as wrong severity ratings.
The Watch · Security desk
What happened
- Each report explains the suspected flaw, includes a proof of concept showing how it could be exploited, and suggests a fix when one exists.
- OSS Scanner is free, modeled on Google's OSS-Fuzz, and uses Anthropic's most capable models to scan enrolled projects periodically.
- Anthropic built the service after some maintainers who triage at scale asked for everything its models had found, unreviewed findings included.
- Projects that stay out keep receiving human-verified disclosures through Anthropic's coordinated vulnerability disclosure process.
- A separate Critical Infrastructure Defense Program has 11 OT security providers as founding partners, among them Dragos, Rockwell Automation and CrowdStrike.
Compiled by The WatchSomething wrong?How this is made
Why it matters
- cost By Anthropic's own forecast, up to one report in ten is wrong, and the maintainer who opted in spends the hours working out which ones.
- decision Maintainers now choose between machine findings at speed and the slower human-checked track, and the choice depends on whether they have the triage staff to absorb the volume.
- constraint In OT, faster discovery adds to a backlog that can take years to clear, because the affected systems often cannot be taken offline to patch.
Anthropic gives the 90% figure as an expectation and says it aims to raise it over time [7]. At that floor, up to about one report in ten would describe a flaw that is not there [17]. Anthropic's own list of inaccuracies to expect includes wrong severity ratings [8]. "The reports are model-generated and sent without human review," Anthropic said [3].
Taking out the reviewer is deliberate. Skipping review gets reports to maintainers faster, according to the company [8]. Anthropic says finding vulnerabilities has never been easier, while verifying, prioritizing and patching them is still hard [9]. Flaws found through its earlier Project Glasswing often took months to fix, and the company admitted Glasswing has not yet cut cyber risk enough [10]. OSS Scanner drops Anthropic's own verification step, one of the hard parts on that list. Projects that say they can keep up with the findings do that work themselves [5].
For operators, the effect shows up in their dependencies. Opt-in projects will hear about flaws sooner [8]. Whether fixes ship sooner depends on each maintainer's triage capacity. The only fix-time evidence Anthropic offered is the Glasswing record of months-long fixes [10]. The announcement as reported does not say how many projects have enrolled or how reports, proofs of concept included, reach maintainers.
Patch cycles in OT are longer still. Anthropic noted that OT systems often cannot be taken offline for patching, so known vulnerabilities can stay open for years. In rare cases, it said, a patch could take decades to apply safely [14]. Through the Critical Infrastructure Defense Program, the companies that handle OT security for power, water, manufacturing and transportation operators get access to Anthropic's frontier Claude models, its threat research and engineers working on site [12]. Several partners already use Claude to fix vulnerabilities and help customers do the same, according to Anthropic [16]. It is starting with a small group and plans to add partners and sectors in the coming months [15].
What to watch
- A measured true-positive rate from Anthropic for OSS Scanner reports, replacing the above-90% expectation.
- Time-to-fix figures from opt-in projects, set against the months-long fixes Anthropic reported for Glasswing findings.
- Whether the CIDP grows past its first 11 partners into the additional sectors Anthropic says it will add in the coming months.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence55
- Adoption25
- Hype gap+10
- Incentives60
- Confidence60
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Anthropic on Thursday announced two new cybersecurity initiatives: one gives open source maintainers faster access to AI-generated vulnerability reports, and the other targets companies that help secure operational technology (OT).
- [2]
OSS Scanner, inspired by Google's OSS-Fuzz, is a free service that uses Anthropic's most capable models to periodically scan open source projects; maintainers have to opt in to have their projects scanned.
- [3]
"The reports are model-generated and sent without human review," Anthropic said.
ReportedSupportedSource: Anthropic, via SecurityWeek2 sources— create a free account to open themView cited source - [4]
Each OSS Scanner report explains the potential vulnerability, includes a PoC showing how it could be exploited and, when one is available, suggests a fix.
- [5]
The service is meant for projects with the capacity to keep up with the findings.
- [6]
Other projects will continue to receive human-verified disclosures through Anthropic's coordinated vulnerability disclosure process.
- [7]
Anthropic expects a true-positive rate above 90% for OSS Scanner reports and aims to improve it over time.
ReportedSupportedSource: Anthropic, via SecurityWeek2 sources— create a free account to open themView cited source - [8]
Skipping review gets reports to maintainers faster, but Anthropic warned that some will contain inaccuracies, such as wrong severity ratings.
ReportedSupportedSource: Anthropic, via SecurityWeek2 sources— create a free account to open themView cited source - [9]
According to Anthropic, finding vulnerabilities has never been easier, but verifying, prioritizing and patching them remains hard.
ReportedSupportedSource: Anthropic, via SecurityWeek2 sources— create a free account to open themView cited source - [10]
Anthropic said Project Glasswing partners uncovered many vulnerabilities but admitted it has not yet cut cyber risk enough; flaws found through Glasswing often took months to get fixed.
ReportedSupportedSource: Anthropic, via SecurityWeek2 sources— create a free account to open themView cited source - [11]
Anthropic launched the service after some OSS maintainers who can triage vulnerabilities at scale asked for everything its AI models had found in their projects, including unreviewed findings.
- [12]
The Critical Infrastructure Defense Program (CIDP) brings frontier Claude models, on-site engineers and Anthropic's threat research to the providers that power, water, manufacturing and transportation operators rely on for OT security.
- [13]
The CIDP's 11 founding partners are Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PwC, and Rockwell Automation.
- [14]
Anthropic noted that OT systems often cannot be taken offline for patching, so known vulnerabilities can remain unresolved for years; in rare cases, it said, a patch could take decades to apply safely.
- [15]
Anthropic is starting the CIDP with a small group of providers and plans to bring the program to more partners and sectors in the coming months.
- [16]
According to Anthropic, several CIDP partners are already working with Claude to fix vulnerabilities and help customers do the same.
- [17]
At Anthropic's expected true-positive floor of 90%, up to about 10 in every 100 OSS Scanner reports would be false positives.
Sources
2 independent publishers whose own reporting we read for this story.
- helpnetsecurity.comAnthropic offers free AI security scans to open-source maintainers
1 article · October 9, 2026
- securityweek.comAnthropic Fast-Tracks AI Bug Reports to OSS Maintainers, Taps 11 Firms for OT Security - SecurityWeek
2 articles · October 9, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.
Topics
Entities
- Nozomi NetworksFollow
- Palo Alto NetworksFollow
- DeloitteFollow
- ClaudeFollow
- Hitachi Ltd.Follow
- OpenSSL CorporationFollow
- OSS ScannerFollow
- Booz Allen HamiltonFollow
- OSS-FuzzFollow
- CrowdStrikeFollow
- Anton ArapovFollow
- DragosFollow
- Project GlasswingFollow
- PwCFollow
- Critical Infrastructure Defense ProgramFollow
- AnthropicFollow
- Insane CyberFollow
- AccentureFollow
- Rockwell AutomationFollow