The 90-day disclosure window has been the security industry’s default for over a decade. Anthropic is now dropping it to 30 days for serious bugs, and the reason is worth paying attention to: AI models are finding the same vulnerabilities independently, and fast.
Anthropic’s Frontier Red Team announced the change to its coordinated vulnerability disclosure (CVD) policy. High- and critical-severity issues now get 30 days. Everything else gets 45. The rules apply to reports sent to vendors and maintainers on or after January 1, 2027, and in-progress reports keep their old deadlines.
What changed
The window is only part of it. Anthropic will also escalate to an external vulnerability coordinator after 15 days of maintainer silence, down from 30. And it will wait 30 days after a patch ships before publishing full technical details, down from 45.
There’s a catch for existing reports. Some high- and critical-severity bugs Anthropic reported more than 90 days ago are still undisclosed. 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. Old CVD norms assumed fixes moved at human speed and that two parties finding the same bug was unlikely. Models now help write fixes, and a bug one model catches is increasingly likely to be caught by another. Anthropic says it is already seeing more rediscovery. If a flaw won’t stay private for 90 days, waiting 90 days mostly protects attackers.
That’s a fair point, and it matters beyond Anthropic. Google Project Zero has long used 90 days, and others have tightened their own timelines. As OpenAI, Google DeepMind and others push vulnerability-finding tools, expect pressure on every vendor’s policy.
The maintainer problem
The obvious objection is that volunteer maintainers can’t fix things in 30 days. Anthropic admits shorter timelines add pressure. Hard cases, like critical infrastructure or bugs needing an RFC change, get a negotiated timeline. Actively exploited bugs keep the seven-day window. AI findings without human review, such as those from OSS Scanner, are exempt.
To offset the burden, Anthropic says it will keep attaching guidance and patches where possible, and agree on a pace before sending large batches. It also points to three programs:
- OSS Scanner, a free service that scans critical open-source projects with its strongest models
- The Claude for Open Source Program, with six months of Claude Ultra at no cost
- The Cyber Verification Program, which adjusts Claude’s safeguards for legitimate defensive work
Does it matter?
Yes, though mostly as a signal. Anthropic is a single reporter, and a shorter clock only helps if maintainers can ship patches. Free scanning and model access are useful, but they won’t fix a staffing shortage. Still, the logic is hard to argue with. If discovery gets cheap and duplicate finds get common, slow disclosure stops being cautious and starts being risky. Other vendors will have to answer the same question.



