Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

AI news & research

Research ·

Anthropic cuts its vulnerability disclosure deadline from 90 days to 30

From 1 January 2027, Anthropic will disclose high- and critical-severity flaws it reports 30 days after notifying maintainers, down from 90, and other issues after 45 days. It will escalate unanswered reports after 15 days instead of 30, and publish full details 30 days after a patch rather than 45. Its reason: AI models now rediscover the same bugs, so they no longer stay private.

Anthropic is giving software maintainers far less time to fix the security flaws it finds. If you maintain an open-source project, or run software that depends on one, the gap between a private bug report and a public write-up is about to shrink by two thirds for the most serious problems.

The new rules apply to reports sent on or after 1 January 2027. High- and critical-severity flaws will be published 30 days after Anthropic notifies the vendor or maintainer, down from 90; everything else gets 45 days. If nobody answers the first report, Anthropic will hand it to an outside vulnerability coordinator after 15 days instead of 30. Once a patch exists, full technical details follow after 30 days rather than 45. Flaws under active attack keep their seven-day window. Older serious reports that are already past 90 days get a final 30 days before they appear on Anthropic's public dashboard.

The company's argument is about rediscovery. Disclosure norms were written when fixes moved at human pace and two researchers rarely found the same bug. Anthropic says it now sees the same flaws turning up again as models improve, so a bug that one model can find should be treated as one that others will soon find. That fits what we have covered from OpenAI, whose pipeline turned up 948 working Linux kernel exploits, and the PHP flaws found independently by two teams. Anthropic's own scanning tools, such as Claude Security, are part of the same shift.

There are exceptions. Bugs in hard-to-patch systems such as critical infrastructure, or flaws that need a change to a specification or RFC, get a negotiated timeline. Findings from Anthropic's automated OSS Scanner that no human has reviewed are outside the policy. Anthropic also offers maintainers of critical projects free scanning, six months of Claude Ultra through its open-source programme, and a verification programme that relaxes Claude's safeguards for defensive security work.

The catch is who carries the load. A month is short for a volunteer team that receives a batch of AI-found bugs, and the company setting the clock is also the one finding the bugs. Anthropic promises patches or mitigations with each report and a sustainable pace for large batches, but has not published how often projects miss the current 90-day deadline. That figure would show whether 30 days is realistic or simply faster.

Our take: The logic is sound, since a bug one model finds another will find too, but the burden lands on open-source maintainers. Anthropic's free scanning and Claude subscriptions help, yet a 30-day clock set by the company finding the bugs is a shift of power worth watching.

Our reviews:Claude Security

Source: Anthropic

More stories