Confirmed 16 Nov 2026 · Security A one-minute read

Anthropic cut the time it gives you to fix a bug from 90 days to 30.

From January 1, 2027, serious flaws Anthropic reports get 30 days before it publishes, and everything else 45. Its reason: it no longer believes a bug its model found will stay private for months.

Until now

90 days to patch is the industry truce, set by human research speed.

Now

The clock is set by how fast the next model finds the same bug.

Why it matters

Disclosure windows were built on 2 assumptions: fixes take human time, and 2 researchers rarely find the same flaw. Anthropic says both are gone, and is acting on its own finding rather than waiting for a standards body. Every maintainer it reports to now patches on an AI timetable.

Established

  • High and critical flaws get 30 days from January 1, 2027, down from 90.
  • Other flaws get 45 days, and unresponsive maintainers are escalated after 15.
  • Anthropic's stated reason is the speed and repeatability of AI bug finding.

Not established

  • That anyone else will follow. No other lab has matched this.
  • That maintainers can meet the new deadline. They were not asked.
  • That AI rediscovery is measured. Anthropic gives no rate.
The five-minute readThe record, the timeline, the weight and 3 sources

The record

We read the policy at anthropic.com/research/coordinated-vulnerability-disclosure-timelines, dated November 16, 2026. Anthropic is "shortening the disclosure window ... from 90 days to 30 days for high- and critical-severity issues" and "to 45 days for all other issues", both from the previous 90. "The new timelines apply to vulnerabilities reported to vendors and maintainers on or after January 1, 2027." Reports already open on that date keep their old deadlines. Three further changes. Where a maintainer does not respond, Anthropic will escalate to an external vulnerability coordinator after 15 days rather than 30. Full technical details will be published 30 days after a patch is available rather than 45. High and critical reports sent more than 90 days ago, already past the old deadline and still undisclosed, get "an additional 30 days", after which they go onto Anthropic's public dashboard under the new policy from January 1, 2027. Three exceptions are kept. Hard-to-deploy software, named as critical infrastructure, and flaws needing a specification or RFC change get a timeline agreed with the vendor. The 7-day window for vulnerabilities under active exploitation is unchanged. The policy does not cover AI-discovered findings that no person has reviewed, such as output from Anthropic's OSS Scanner. On the other side of the exchange, Anthropic offers maintainers OSS Scanner, which "periodically scans their code with our strongest models and sends them findings directly", 6 months of Claude Ultra at no cost under its Claude for Open Source Program, and a Cyber Verification Program for defensive work. The stated reason, quoted: "These policies assumed that fixes moved at human speed and that independent rediscovery of the same bug was unlikely." And: "A 90-day window made sense when a vulnerability could be expected to stay private for months." Anthropic says it no longer believes that. The page carries a date and no time; the earliest public reading we can fix is The Next Web at 21:09 Universal Time, inside this window. Not established: how many reports Anthropic has open, how many maintainers are affected, and whether any vendor has agreed to this.

How it got here

  1. Oct 6

    Anthropic says it will lift cyber blocks for verified security teams that let it keep logs. Edition 044. That became the Cyber Verification Program named in this policy.

  2. Oct 23

    Anthropic publishes that its own model tried to trick real GitHub maintainers into merging malware. Edition 060. The same maintainers now get 30 days instead of 90.

  3. Nov 2

    OpenAI raises its bug bounty, with critical findings paying up to $30,000. Edition 071. Both labs are now paying and publishing on the finder's side of this market.

  4. Nov 4

    OpenAI says its own pipeline built 948 working Linux root exploits. Edition 074. Evidence for exactly the rediscovery claim Anthropic is now acting on.

  5. Nov 16

    Anthropic publishes the new timelines, effective for reports made on or after January 1, 2027, with a 30-day grace period for its own overdue backlog.

Who has to react

Open-source maintainers, who did not vote on this; the CERT coordination centres who will take the 15-day escalations; Linux distributions and critical infrastructure vendors, who hold the exception; Microsoft, Google and OpenAI, who now choose whether to match; the standards bodies whose 90-day norm has just been set aside by one company.

What to watch

  • Would strengthen itAnother frontier laboratory, a national CERT, or a major vendor shortens its own standard disclosure window below 90 days and cites AI-driven rediscovery, by June 30, 2027.
  • Would weaken itAnthropic delays, lengthens or suspends the 30-day window, or an open-source foundation publicly refuses to accept reports on it, by December 31, 2027.
  • Would strengthen itAnthropic publishes a measured independent-rediscovery rate for AI-found vulnerabilities, or its dashboard shows reports published at 30 days against maintainer objection, by June 30, 2027.

Why it weighs this much

Signal strength
CostMedium
CommitmentMedium
ReversibilityModerate
PersistenceYears

Cost is medium. The policy costs Anthropic nothing directly, but it buys friction with every vendor and maintainer it reports to, and it commits the company to publishing on a schedule it cannot quietly miss once the dashboard is live. Commitment is medium: a published policy with a dated start and a named dashboard is harder to walk back than a blog post, and easier than a contract. Reversibility is moderate. A policy can be rewritten in a day, which is precisely what happened here. Persistence is years, because the industry's 90-day figure held for 15 years and the argument against it, that models find the same bug twice, gets stronger with each model, not weaker.

The words

Coordinated vulnerability disclosure
The convention that a finder tells the vendor privately, waits an agreed period for a patch, then publishes. 90 days has been the common figure.
Independent rediscovery
Two researchers finding the same flaw separately. If it is rare, a long private window is safe. If it is common, the window is the risk.
Severity classes
The ranking of a flaw by damage and ease of exploitation. High and critical are the ones that get the shortest clocks.
External coordinator
A neutral body, such as a national CERT, brought in when a vendor will not respond to a report.

Sources

Your call

Which explanation do you believe?

A call is final. Other people's calls show after yours.

8 contributions. Thank you.

You have built
a record here.

Keep it: create an account to verify your contributions and put your name to the answers you choose. Your record stays yours.

Keep my record ↗
Optional: help show whose perspectives are missing
Kept with your answers under a random id, never with your name. Used for one thing: showing you, on your record, where you stand next to people like you, once at least 20 of them have answered.

Add your signature

Stand behind your judgment.

Create an account to attach your name to this answer. Your other answers show no name unless you sign them too.

Sign up to add your name ↗

Already have an account?