Samuel Arendt

Anthropic cuts the disclosure deadline for serious flaws found by Claude from 90 to 30 days

On November 16, 2026, Anthropic's Frontier Red Team announced shorter deadlines in its coordinated vulnerability disclosure policy, the rules the company follows when Claude finds a flaw in third-party software. For high- and critical-severity flaws, the time between notifying the maintainer and public disclosure drops from 90 to 30 days. For everything else, it drops to 45 days. The new deadlines apply to flaws reported on or after January 1, 2027; reports already in progress on that date keep the old deadline.

The three deadlines that change

The current policy, in the version published in March 2026, follows the industry's 90-day standard, with a possible 14-day extension when the maintainer is working on a fix. The announcement changes the disclosure window and two related deadlines:

Step

Today

From Jan 1, 2027

Disclosure of a high or critical flaw

90 days

30 days

Disclosure of other flaws

90 days

45 days

Escalation to an external coordinator when the maintainer doesn't respond

30 days

15 days

Publication of full technical details after the patch

45 days

30 days

The seven-day deadline for actively exploited flaws stays the same. Anthropic makes an exception for hard-to-deploy software, such as critical infrastructure, and for flaws that require changing a specification or RFC; in those cases it says it will agree on a timeline with the vendor. AI findings without human review, such as those from the OSS Scanner service, are outside the change.

The company's reasoning is that models have become much better at finding flaws and that the same flaw is now rediscovered more often. In the announcement's words, a 90-day window made sense when a vulnerability could be expected to stay private for months, and the company no longer believes it can.

There is also a backlog. High and critical flaws reported more than 90 days ago, already past the current policy's deadline and still undisclosed, get 30 more days from the notice Anthropic says it is now sending to maintainers. After that, the reports go to the company's public disclosure dashboard.

Who feels the change

The policy covers flaws Anthropic finds in open-source software and in closed-source software when it has permission to do the research. The effect arrives from two directions.

The first is for anyone who maintains public code: a company that publishes a package on Packagist or npm, a WordPress plugin, a library its customers use. If a high-severity report arrives in 2027, the team will have 30 days to confirm, fix and ship the release. Anthropic itself acknowledges this puts more pressure on maintainers and says it will keep sending practical guidance, and a patch or mitigation where it can, and will agree on a pace before sending many findings to the same project.

The second is for teams that only use those dependencies, which is almost every web product team. Once a patch is out, the full technical details become public in 30 days instead of 45. That is the window a team has to update before anyone can read a description of how to exploit the flaw. Keep in mind this is Anthropic's rule; other researchers and companies follow their own policies.

What to do before January

If you maintain a public project, the work is making sure a report reaches someone and gets an answer in under 15 days: a security address someone actually reads, private vulnerability reporting turned on in the repository, and a person with time set aside for triage. Without an answer in 15 days, the case goes to an external coordinator. It also helps to know, before the first report, how long the project takes today to ship a fix, because 30 days includes review and release.

If you consume dependencies, the question is how long it takes from a patch being published to it reaching production. If the answer is more than 30 days because of a deploy queue, manual staging sign-off, or a runtime version that blocks updates, that is the point to fix. I wrote about why a supported version still calls for a migration decision; shorter disclosure deadlines make that decision more expensive to postpone.

If you don't maintain public code and already update dependencies within a few weeks, there is nothing to change right now.

What is still unclear

The policy page, checked on November 20, still shows the 90-day deadline and the March update date; the revised text has not been published. The announcement says the rest of the policy stays the same, but doesn't confirm whether the 14-day extension still applies under the new 30-day deadline.

It is also unknown how many reports are in the overdue backlog or which projects they affect. The announcement's wording leaves open whether those reports will be published at the end of the extra 30 days or only from January 1. And Anthropic speaks only for itself: the announcement doesn't say whether other companies using AI to find flaws will adopt similar deadlines.

If your company is weighing something similar, reach me on LinkedIn or WhatsApp.

References

Empty concrete staircase seen from below, with parallel metal handrails running up the middle of the steps

Photo by Lucas Davies on Unsplash

#AI Agents, #Operations, #Radar