Anthropic is offering open source maintainers free, recurring security audits from its most advanced frontier AI models. But experts warn the service could exacerbate a problem already straining under-resourced projects — AI is finding potential vulnerabilities much faster than humans can review them.
"The real measure of success shouldn't be the number of vulnerabilities discovered but the number that are validated, prioritized and remediated before attackers exploit them," Paul Nashawaty, analyst at TheCube Research, told TechTarget.
Anthropic's OSS Scanner — an opt-in security vulnerability scanner for eligible open source software projects — opened enrollment on Oct. 8. Participants will receive free, recurring code reviews by the AI lab's most powerful models, including Claude Mythos.
According to Anthropic, each report includes a self-contained reproducer, an explanation of the vulnerability and, if available, a candidate patch. But because outputs are completely model-generated, without human review or triage — enabling participants to receive scan results faster and more frequently — Anthropic said reports could be incorrect or invalid.
Many open source advocates greeted OSS Scanner with enthusiasm, saying it will give them insight into possible flaws in their code that they could not otherwise access. Some experts warn, however, that the service offloads the significant operational burdens of manual validation and triage to maintainers — and could easily overwhelm small teams.
"The scan is free, but the time needed to deal with its findings isn't," said Krishna Shakula, an AI engineer at the law firm Robertson, Anschutz, Schneid, Crane & Partners. "Someone still has to reproduce the issue, check whether it's actually exploitable in their project, decide how urgent it is and review the proposed fix. Even a real vulnerability can take substantial work to resolve safely."
Anthropic's AI finds possible flaws faster than humans can review them
As of Oct. 2, Anthropic said its advanced models, including Claude Mythos, had found 29,432 candidate vulnerabilities in open source software. Notably, however, the frontier AI lab was able to manually review and triage only 6,123 of those, ultimately resulting in 513 findings patched upstream.
The scan is free, but the time needed to deal with its findings isn't.
Krishna Shakula
AI engineer, Robertson, Anschutz, Schneid, Crane & Partners
"While 6,000 vulnerabilities is significant, we remain bottlenecked on our human capacity to validate these findings," Anthropic said in an online statement, adding that open source maintainers have increasingly asked to receive bulk submissions of all unverified reports and proposed patches. That demand for unvalidated findings reportedly drove Anthropic to create the free OSS Scanner service.
"Since exploits can now be developed in minutes, projects that can find and address vulnerabilities faster can better secure their software against attackers who are racing to find and leverage these same weaknesses," Anthropic said.
According to the AI lab, when pen testers checked 97 critical and high-severity vulnerabilities from OSS Scanner across 48 projects, they found 88% met the bar for disclosure. Eleven were duplicate findings and one was a false positive. Some maintainers reported that the scanner inflated severity ratings, Anthropic said, but "seldom" that a critical or high-severity finding was invalid.
Open source is ''the weakest link in enterprise software supply chains''
At the OpenSSL Corporation, which oversees the open source cryptographic library that secures much of the internet's encrypted traffic, external vulnerability reports have skyrocketed. And according to Director Anton Arapov, the OpenSSL security team manually reviews and triages each one.
"Verification is still done by people," Arapov wrote on LinkedIn. "We had to grow the [security] team to keep up."
Unreviewed candidate vulnerabilities — such as the roughly 23,000 that Anthropic's models found that went unvalidated due to the review bottleneck — leave projects like OpenSSL at risk, he added.
"Findings that reach the maintainer get fixed once, upstream, for everyone. Findings that stay with whoever found them get patched in private forks," Arapov wrote. "The OpenSSL Library will enroll in OSS Scanner for that reason."
Nashawaty said Anthropic's OSS Scanner is a potentially important step toward addressing what he called "the weakest link in enterprise software supply chains" — open source software. The economics of the service are compelling, he added, with maintainers able to access findings from gated frontier models for free.
"OSS Scanner could strengthen the supply chain, but only if the industry invests in the complementary capabilities such as automated triage, patch validation, maintainer funding and enterprise participation that turn AI-generated findings into fixes," Nashawaty said. "Otherwise, we risk automating vulnerability discovery while leaving the remediation bottleneck untouched."
OSS Scanner could help — or hurt — open source maintainers
Shakula expressed cautious optimism about Anthropic's OSS Scanner program, while also noting that it won't be right for everyone.
Consider, for example, a small open source project with just one person managing everything from releases to support. Before enrolling in Anthropic's service, Shakula recommended that projects clearly establish who will review the OSS Scanner reports and how much time they can reasonably dedicate to it. If the team already struggles to manage confirmed security issues, another stream of findings could make prioritization even more challenging.
"Getting a real issue earlier has value. But faster delivery alone doesn't tell us whether the overall workload improved," Shakula said, adding that he would measure success as confirmed vulnerabilities fixed per maintainer-hour. "I'd want to know how much maintainer time it takes to turn those findings into safe fixes."
If a project has the capacity to independently validate AI-discovered candidate bugs, then enrollment in Anthropic's OSS Scanner could be a significant boon, according to Shakula. He noted that participants can choose to pause reports, and unvalidated findings don't currently come with forced disclosure deadlines.
"A project with capacity to investigate could benefit a lot," he said. "A stretched, volunteer maintainer may reasonably decide to wait. Both are sensible decisions."
Alissa Irei is an Informa TechTarget senior reporter covering cybersecurity news.
