Anthropic cuts its vulnerability disclosure window to 30 days, and the reason should worry you

Ninety days was the industry’s comfortable fiction. Anthropic is now saying out loud that it no longer holds, because the same models that find bugs for defenders will find them for everyone else too.

The company announced that its coordinated vulnerability disclosure (CVD) policy will drop from 90 days to 30 days for high- and critical-severity issues, and to 45 days for everything else. The new clock applies to reports sent to vendors and maintainers on or after January 1, 2027. Reports already in progress keep their old deadlines.

What else is changing

The window is not the only thing that moved. Anthropic will now escalate to an external vulnerability coordinator after 15 days of maintainer silence, down from 30. It will also wait 30 days after a patch ships before publishing full technical details, instead of 45. The seven-day window for bugs under active exploitation stays as is.

And there is a catch for existing reports. Some high- and critical-severity bugs Anthropic reported more than 90 days ago are still not public. Maintainers are being told they get 30 more days, then the reports go on Anthropic’s dashboard.

Why the old assumptions broke

Anthropic’s argument is simple. Disclosure norms assumed fixes moved at human speed and that two parties independently finding the same bug was unlikely. The company says neither is true now. Models can help write patches, and a flaw one model finds is increasingly likely to be found by another. Anthropic says it is already seeing more rediscovery.

That is the part users should care about. If a bug can be rediscovered quickly, a long private window does not keep it secret. It just leaves you exposed while you assume it is. A quiet 90 days stops being a safety margin.

The cost lands on maintainers

But shorter deadlines are not free. Thirty days is tight for a volunteer maintaining a widely used library. Anthropic concedes this. Its mitigations are guidance with each report, patches or workarounds where possible, and agreeing on a pace before sending a large batch to one project. Hard cases, like critical infrastructure software or flaws needing a spec or RFC change, get negotiated timelines.

Anthropic also points to programs for the people on the receiving end:

  • OSS Scanner, a free service that scans critical open-source projects with its strongest models and sends findings directly.
  • The Claude for Open Source Program, offering eligible maintainers six months of Claude Ultra at no cost.
  • The Cyber Verification Program, which adjusts Claude’s safeguards for legitimate defensive work.

My take

The logic is sound, but note the incentives. Anthropic sells the tools that create the flood of findings, then offers more of its own tools to cope with it. Also, the policy exempts AI-discovered findings that have not received human review, such as OSS Scanner output. That is sensible, but it means the firehose of unreviewed reports sits outside the deadline.

Expect other labs and researchers to face the same pressure. If disclosure timelines shrink across the industry, patch capacity, not discovery, becomes the bottleneck. That is where users’ privacy will actually be won or lost.