AI changes the economics of rediscovery
Anthropic is shortening its coordinated vulnerability disclosure policy from a standard 90-day window to 30 days for high and critical issues and 45 days for other severities. The new deadlines apply to reports sent on or after 1 January 2027, while existing cases keep their current timetable. Anthropic says frontier models are discovering vulnerabilities faster and that independent rediscovery is becoming more likely, reducing the period during which a privately reported flaw can reasonably be expected to remain unknown to attackers.
The policy change is a practical consequence of capability growth rather than a new model launch. If several researchers and automated systems can find the same flaw, a long private window may leave users exposed after adversaries have independently reached the result. Shorter deadlines increase pressure on maintainers, especially open-source projects with limited staff, but they can also create a clearer incentive for vendors to acknowledge and prioritise a serious report. The challenge is preserving quality and coordination while moving faster.
For high and critical findings, the deadline drops by two-thirds. Other vulnerabilities receive 45 days. Anthropic’s seven-day window for flaws under active exploitation remains unchanged. Severity therefore becomes even more consequential. Reports should explain the scoring method, affected versions, realistic attack path and environmental assumptions so maintainers can challenge a classification. Inflated severity can force a disruptive response, while understated risk delays protection for users who may already be exposed.
High-severity deadlines fall to 30 days
Anthropic will aim to involve an external vulnerability coordinator after 15 days without a maintainer response, down from 30. A coordinator can help confirm contact details, reach a vendor and balance disclosure where ownership is unclear. This is particularly useful for abandoned dependencies and projects maintained by volunteers. Escalation should not automatically imply public blame. Researchers need to document contact attempts and give maintainers a secure, practical channel to receive evidence and proposed mitigations.
After a patch becomes available, Anthropic plans to wait 30 days rather than 45 before publishing full technical detail. That interval allows users to update while preserving the educational and accountability benefits of disclosure. In ecosystems with slow downstream packaging, 30 days may still be tight. Advisories should identify fixed versions and practical mitigations early, while detailed exploit mechanics can wait. Vendors and distributors need automation that maps upstream fixes into the products and fleets that actually depend on them.
Maintainers need evidence and sustainable volume
Some vulnerabilities cannot be resolved on the standard clock. Critical infrastructure, widely implemented specifications and difficult deployment environments may require architectural change or coordinated updates across several organisations. Anthropic says it will work with vendors on alternative remediation timelines in those cases. Flexibility is important, but exceptions need documented milestones and interim protections. An open-ended extension recreates the uncertainty the new policy is trying to remove.
Anthropic also addresses a backlog of high and critical reports that already exceed 90 days. Maintainers will receive an additional 30 days of notice before reports are published through the company’s dashboard under the updated policy. That is a material operational event for affected projects. Maintainers should verify contact channels, review dormant security inboxes and prepare advisories rather than discovering a disclosure through public coverage. Users of critical open-source components should monitor the same dashboard and vendor notices.
Human review remains a publication boundary
The policy explicitly excludes AI-discovered findings that have not received human review, including raw results from Anthropic’s OSS Scanner. That boundary is essential. Automated scanners can generate plausible but invalid attack paths, duplicates and findings that ignore deployment context. Human validation should reproduce the issue, assess impact and remove unsafe or irrelevant details before a disclosure clock begins. Volume must also be paced so a small project is not overwhelmed by a burst of machine-generated reports.
Anthropic’s decision signals that disclosure norms are being recalibrated for AI-assisted security research. Faster discovery can help defenders only if patches, distribution and user updates accelerate too. Australian software teams should inventory their public security contacts, define severity and escalation ownership and rehearse a 30-day high-risk remediation. Maintainers also need support: actionable reproduction, a patch where possible and a sustainable reporting cadence. The goal should be shorter exposure, not faster publication for its own sake.
The shorter clock also changes how product teams should handle inbound reports. Acknowledgement, duplicate checking, reproduction and owner assignment can no longer consume most of the first month. Organisations need an on-call security intake, a maintained component inventory and authority to ship an emergency fix outside the normal release train. They should measure time to first response and time to a deployable mitigation, not simply whether publication was delayed. Suppliers should include downstream customers and managed-service partners in that plan, since a patch that exists but cannot be identified or deployed leaves the practical exposure unchanged.