
Anthropic is cutting the time it normally gives software makers to address serious security flaws it discovers before public disclosure. Under an updated policy announced on 16 November, its window for high- and critical-severity vulnerabilities will fall from 90 days to 30. Other flaws will have a 45-day window. The new timetable applies to reports sent to vendors and maintainers from 1 January 2027, not to cases already in progress on that date.
The change reflects a difficult consequence of AI-assisted security research. Models can help defenders find and fix bugs, but Anthropic argues that they also make it more likely that someone else will independently find the same weakness before a conventional 90-day period has expired. Shorter deadlines aim to reduce the interval during which a privately known flaw could be rediscovered and misused.
Anthropic says it will seek help from an external vulnerability coordinator after 15 days without a response from a maintainer, down from 30 days. Once a patch is available, it plans to wait 30 rather than 45 days before publishing full technical details. Its existing seven-day window for vulnerabilities under active exploitation remains unchanged.
For high- and critical-severity findings that are already beyond the existing 90-day deadline but remain private, Anthropic says it will notify maintainers and give them another 30 days before adding the cases to its disclosure dashboard from January. That is a transition measure, not an immediate publication of every unresolved report.
The policy does not automatically apply to raw AI-generated findings. Anthropic says issues must receive human review, while findings from its OSS Scanner that have not been reviewed fall outside this update. This distinction matters because an automated scan can produce leads that need validation before a vendor is asked to treat them as genuine vulnerabilities.
Thirty days may be enough for a contained fix in a well-resourced project. It can be much harder for maintainers of widely deployed open-source software, industrial systems or products requiring specification changes. A patch may need testing, coordinated releases and time for downstream users to deploy it. Public disclosure before that work is complete can increase exposure, even when the researcher intends to improve security.
Anthropic acknowledges that some flaws cannot be resolved on a fixed schedule and says it will work with vendors on an appropriate remediation timeline in such cases. It also promises actionable reports, suggested fixes or mitigations where possible, and a sustainable pace for maintainers handling many findings.
The policy illustrates a broader shift in cybersecurity. AI does not simply increase the number of bugs researchers may find; it changes how quickly discovery, verification and patching must be coordinated. Whether Anthropic’s shorter windows help users will depend on the quality of its reports, the support given to maintainers and how flexibly difficult fixes are handled.







