Anthropic Cuts Vulnerability Disclosure Window From 90 to 30 Days
From 2027, maintainers will get 30 days to fix high- and critical-severity bugs found by Claude before details go public, as Anthropic says AI makes the same flaws easier for others to rediscover.
Published 5 min read

Anthropic will give software maintainers far less time to fix the security bugs its Claude models find before it makes them public. In a policy update published on 2026-11-16, the company said it is cutting its standard disclosure window from 90 days to 30 days for high- and critical-severity issues, and to 45 days for everything else.
The new deadlines apply to vulnerabilities reported on or after 1 January 2027 (2027-01-01). Reports already open on that date keep their current deadlines. Anthropic also plans to escalate faster when maintainers stay silent and to publish full technical details sooner after a patch ships.
What changes in Anthropic's vulnerability disclosure policy
Coordinated vulnerability disclosure (CVD) is the usual way security researchers handle a bug: they report it privately to the vendor or open-source maintainer, allow a fixed period to ship a fix, and then publish the details so defenders can act. The length of that period is the main thing Anthropic is changing.
| Rule | Current policy | From 2027-01-01 |
|---|---|---|
| Disclosure window, high and critical severity | 90 days | 30 days |
| Disclosure window, other issues | 90 days | 45 days |
| Escalation to an external coordinator if the maintainer does not reply | After 30 days | After 15 days |
| Wait after a patch before full technical details | 45 days | 30 days |
| Bugs under active exploitation | 7 days | 7 days (unchanged) |
Anthropic said the rest of its current CVD policy stays the same. That page, last updated in March 2026, calls 90 days "the industry standard" deadline and allows a 14-day extension when a maintainer is engaged and close to a fix. It also says Anthropic generally defers to the maintainer when the two sides disagree on how severe a bug is.
Why Anthropic says 90 days no longer works
According to Anthropic's Frontier Red Team, frontier AI models have become much better at discovering vulnerabilities over the past several months. Older disclosure norms, the company argues, assumed that fixes moved at human speed and that nobody else was likely to find the same bug independently. "Neither assumption holds today," Anthropic wrote.
The company said models can now help write fixes, and that "a bug that one model can find is increasingly likely to be found by another." It reports seeing a higher rate of rediscovery, where the same flaw is found separately by different parties. "A 90-day window made sense when a vulnerability could be expected to stay private for months. We no longer believe it can," Anthropic said.
The concern is not theoretical. Google's threat intelligence team reported this week that attackers had used an AI framework to turn security patches into working exploits with no human in the loop, which shrinks the safe gap between a fix and its abuse.
The numbers behind the pressure on maintainers
Anthropic publishes a disclosure dashboard for bugs found by Claude models, including an early snapshot of Claude Mythos Preview, and validated with outside security research firms. As of 2026-11-02, it listed 5,087 vulnerabilities disclosed across 680 open-source projects, of which 629 had been patched to Anthropic's knowledge. Those findings had produced 733 identifiers: 230 CVE records and 503 GitHub Security Advisories.
By our calculation from those figures, about 12% of the disclosed bugs had a known patch at that point. Anthropic notes that some maintainers ship fixes without publishing an advisory, so the real share may be higher. Even so, the gap shows how much work a shorter clock puts on maintainers.
The dashboard also shows that the reported bugs are only part of what the models find. For findings discovered between 2025-11-01 and 2026-11-02, Anthropic triaged 37,966 raw findings, narrowed them to 6,376 candidates and had 6,009 reviewed by external firms, with a 94.2% true-positive rate. The company says human triage is the step that limits how many bugs it can report.
What happens to bugs that are already overdue
Some high- and critical-severity bugs that Anthropic reported more than 90 days ago have already passed the deadline in its current policy but are still not public. Anthropic says it is telling those maintainers they have an extra 30 days, after which it will publish the reports on its dashboard under the updated policy starting 2027-01-01.
Exceptions and help for open-source projects
Anthropic accepts that not every flaw can be fixed in 30 or 45 days. Bugs in hard-to-deploy software such as critical infrastructure, and flaws that need a change to a specification or an RFC (the documents that define internet standards), will get a timeline agreed with the vendor. The shorter windows also do not apply to AI-discovered findings that have not been reviewed by a human, such as results from its automated OSS Scanner service.
To soften the impact, Anthropic says each report will keep including practical guidance and, where possible, a patch or mitigation, and that it will agree a sustainable pace before sending a large batch of findings to a single project. It points maintainers to three programs:
- OSS Scanner: free periodic scans of critical open-source projects with Anthropic's strongest models, with findings sent straight to the maintainers.
- Claude for Open Source Program: six months of Claude Ultra at no cost for eligible maintainers and contributors.
- Cyber Verification Program: adjusted Claude safeguards for security professionals doing legitimate defensive work.
For administrators, the practical takeaway is that, for bugs Anthropic reports, the time between a fix appearing and full technical details going public will get shorter. Long-lived bugs such as the 11-year-old NoisyTCP Linux kernel flaw show how long a flaw can sit in widely used code, so teams that depend on open-source projects may want to apply updates faster from 2027.