Finding software vulnerabilities is no longer the slow part. Fixing them is. Anthropic's disclosure dashboard shows 6,157 findings reported to open-source maintainers as of October 2, 2026, but only 516 patched upstream, to the company's knowledge. On October 8, Anthropic launched its Cyber Mission to close that gap, and it matters to anyone running AI tools with a real token budget behind them.
Here are the numbers, the two programs, and what security teams should do now.
The Disclosure Numbers
| Metric (Nov 1, 2025 to Oct 2, 2026) | Figure |
|---|---|
| Findings discovered | 29,439 |
| Reported to maintainers | 6,157 across 591 open-source projects |
| Reviewed by outside firms | 6,123 reviewed, 5,674 confirmed valid (about 92.7%) |
| Reports that went through triage | 1,333 |
| Sent directly to maintainers | 4,824 (may contain false positives) |
| Patched upstream | 516 (about 8.4% of reported) |
Progress is real but uneven. Between May 22 and October 2, reported findings grew from 1,596 to 6,157 (3.9x), while upstream patches grew from 97 to 516 (5.3x). Patches are gaining, but they remain a small slice of the total.
The Two Tracks of the Cyber Mission
1. Critical Infrastructure Defense Program
Anthropic sends frontier Claude models, on-site engineers, and threat research to 11 partners: Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PwC, and Rockwell Automation. Operators reach this help through their providers, not directly. Axios reported that Anthropic did not say whether partners get free model access or who pays for computing.
2. OSS Scanner
Enrolled open-source projects get free periodic scans, each with a reproducer, an explanation, and a candidate patch when one exists. Anthropic states that these reports are model-generated and sent without human review, though it will keep human-reviewed disclosure for projects that want it. Maintainers enroll through a pull request.
Anthropic also says security teams of any size can apply to its expanded Cyber Verification Program, and that its June program for state, local, tribal, and territorial governments reached more than half of US states.
Scanner Accuracy: Early Results
- Penetration-tester check: Of 97 critical and high findings in 48 projects, 85 met the disclosure bar, 11 were real but duplicated known issues, and one was a false positive.
- Maintainer feedback: wolfSSL reported that all but two of its 74 reports were valid and five became CVEs. Other maintainers said severity ratings can be inflated or the project's threat model misunderstood.
Why Patching Is the Bottleneck
In Anthropic's earlier Glasswing work, months often passed between a vulnerability being found and fixed. In operational technology (OT), a fix may have to wait until it is safe to apply to running machinery, and Anthropic says that in rare cases this could take decades.
The risk is current. On July 30, the FBI and EPA warned that water utilities in at least seven states had reported attacks on internet-facing Rockwell Automation Allen-Bradley MicroLogix 1100 and 1400 controllers since July 27, with effects including loss of pressure and flooding. Their first recommendation was to remove PLCs from direct internet exposure using secure gateways and firewalls. CrowdStrike's Adam Meyers said in April that "the problem is patching," warning that at ten times today's findings, teams would be "completely underwater."
Five Checks for Security Leaders
- Check provider access: Ask your OT provider whether it is one of the 11 partners and what your organization actually receives.
- Remove PLC internet exposure: Work with your OT team to take PLCs off direct internet exposure. This is a risk today, whatever happens with fix timing.
- Track open-source dependencies: Reports go to maintainers, not to you, so watch advisories and upstream releases for projects you depend on.
- Verify fixes are installed: Track patch release and patch installation separately, and log stopgap protections where patching must wait for a maintenance window.
- Apply for access directly: If your team does defensive work, apply to the Cyber Verification Program yourself. A provider's seat does not pass model access through to you.
Why This Matters
AI is making vulnerability discovery cheap and fast, but the remediation chain of maintainers, vendors, and operators still moves at human speed. The risk is a growing backlog of known flaws, especially in infrastructure that cannot be patched quickly. Anthropic's numbers show the gap narrowing slowly, not closing.
Official sources: Anthropic Cyber Mission announcement and the coordinated disclosure dashboard.
Frequently Asked Questions (FAQ)
What is the Anthropic Cyber Mission?
It is an initiative announced October 8, 2026, with two tracks: the Critical Infrastructure Defense Program for 11 provider partners, and OSS Scanner, a free opt-in scanning service for open-source projects.
How many vulnerabilities have been patched?
As of October 2, 516 of 6,157 reported findings were patched upstream, to Anthropic's knowledge. That does not mean the patches are installed everywhere.
Are the AI-generated findings accurate?
Outside firms confirmed 5,674 of 6,123 reviewed findings as valid. But 4,824 reports went directly to maintainers and may contain false positives.
Can my organization get access?
Security teams of any size can apply to the Cyber Verification Program, and Anthropic verifies applicants before granting access. Open-source maintainers can enroll in OSS Scanner by pull request.
Who pays for model access in the infrastructure program?
Anthropic has not published that. A proposed $100 million CISA pilot would fund free access for operators, but only if Congress appropriates the money.
No comments:
Post a Comment