Back to Explore
NewsVibeFix editorialUpdated Nov 20, 2026

Anthropic cuts the vulnerability disclosure clock from 90 days to 30: AI finds bugs too fast for the old rule

From 1 January 2027, Anthropic will disclose high- and critical-severity vulnerabilities after 30 days instead of 90, and all other issues after 45. Google's Project Zero moved the same week. The reason in both cases: AI has made the old 90-day assumption obsolete.

A security operations center with a wall of large monitors showing world threat maps and data charts

In short: Anthropic cuts the countdown to publishing a vulnerability from 90 days to 30, effective 1 January 2027

On 16 November 2026, Anthropic announced on its official research blog that it is updating its coordinated vulnerability disclosure (CVD) policy timelines: from 1 January 2027, for vulnerabilities reported to vendors and maintainers on or after that date, the disclosure window for high- and critical-severity issues drops from 90 days to 30 days, and for all other issues from 90 days to 45 days. If a maintainer does not respond, Anthropic will escalate to an external vulnerability coordinator after 15 days instead of 30; after a patch is available, the wait before publishing full technical details drops from 45 days to 30 days. The one thing that does not change is the 7-day window for vulnerabilities already under active exploitation — those still go public in 7 days.

This is not Anthropic acting alone. Three days earlier, on 13 November, Google's Project Zero announced a similar overhaul: scrapping the "90 days to fix plus 30 days of soak time" policy it had run since 2021, and replacing it with a fixed number of days from the report date — 90 days now, 60 days from 2 February 2027, and 30 days from 3 January 2028. Two teams moving in the same direction in the same week, for the same reason: AI has made finding and exploiting vulnerabilities faster, and the old 90-day assumption no longer holds. The news here is Anthropic's policy; Project Zero's change is included as context, because only together do the two show this is an industry-wide shift in the rules, not one company's whim.

A note on sourcing, first. Anthropic's official page carries its own date stamp — Published 2026-11-16, updated 2026-11-16T14:32:08Z — so it is a primary source. Project Zero's 13 November policy comes from that team's own blog, and this article cross-checked its date and key figures through several independent reports (BleepingComputer, SecurityWeek and others) rather than quoting the Project Zero post word for word. The dashboard statistics cited here are as of 2 November 2026, and that date is stated wherever the numbers appear.

Old versus new, side by side: four things shorter, one unchanged

Start with the timelines side by side. The disclosure window is the core: previously a flat 90 days regardless of severity; under the new rules it splits into two tiers, 30 days for high- and critical-severity issues and 45 days for everything else. A critical vulnerability's "quiet period" loses 60 days at a stroke; a medium- or low-severity one loses 45. Anthropic's explanation is that the 90-day window was designed around the idea that a vulnerability could be expected to stay private for months — and that idea no longer holds, so the window has to shrink with it.

The second change is "what if nobody answers". If a maintainer does not respond to the initial report, Anthropic used to wait 30 days before escalating to an external vulnerability coordinator; now it waits 15. The coordinator acts as a go-between — helping contact the maintainer, confirming the bug and pushing the fix along. For a small project with one or two maintainers who reply slowly, that 15-day difference can decide whether a vulnerability gets fixed or sinks without a trace.

The third change is the waiting period after a patch ships. Previously, Anthropic waited 45 days after a patch was available before publishing full technical details, giving users time to install it; now it waits 30 days. The logic is the same as Project Zero's: the patch itself leaks details of the vulnerability, and an attacker with the patch can reverse-engineer the bug, so the "soak time" does not stop a determined attacker — it only slows defenders from getting the full picture.

The fourth item is not a rule change but a backlog clean-up. Anthropic says there is a set of high- and critical-severity vulnerabilities reported more than 90 days ago, already past their deadline under the current policy, that have still not been disclosed publicly. The maintainers of those projects are now being notified and given a further 30 days, after which the reports will be published on Anthropic's dashboard in line with the updated policy, starting 1 January 2027. The number of reports in this backlog was not disclosed, but they will be the first things published once the new policy takes effect — maintainers of affected projects should check the dashboard to see whether they are on the list.

The one thing that stays put is the 7-day window: vulnerabilities already being exploited in the wild still go public in 7 days. That window was the shortest from the start, and Anthropic has not touched it — a sign that on "already under attack" there was never much room for argument.

Anthropic's coordinated vulnerability disclosure policy, old vs new: high- and critical-severity issues 90 days down to 30, all other issues 90 down to 45, escalation for non-responsive maintainers 30 down to 15, full technical details after a patch 45 down to 30, vulnerabilities under active exploitation unchanged at 7 days

Why now: AI rewrote the assumption that "a bug can stay private"

To understand this change, you have to understand what the old rule was designed around. The core bargain of coordinated disclosure is: the finder tells the vendor privately first, the vendor gets time to fix, and the details go public when the time is up. That system rested on two assumptions — first, that fixing moves at human speed, and 90 days is enough for a team to write, test and ship a patch; second, that the chance of someone else independently rediscovering the same bug is low, so during the quiet period the finder is probably the only one who knows. Anthropic says plainly in its blog post that neither assumption holds any more.

Take rediscovery first. Anthropic's own words: a bug that one model can find is increasingly likely to be found by another. They are seeing a higher rate of vulnerability rediscovery as these capabilities increase. In other words, your vulnerability report is no longer "exclusive intelligence" — keeping it quiet for 90 days does not mean attackers are in the dark for 90 days too. The longer the quiet period, the smaller the information gap between defenders and attackers may become: defenders are waiting for a patch while attackers may already be scanning out the same bug with their own models.

Then fixing speed. Models do not only find vulnerabilities; they can help write the fixes, too. Anthropic's argument is that since discovery and fixing have both sped up, the disclosure rhythm should keep pace. Project Zero says the same thing: in its 13 November post, Natalie Silvanovich wrote that AI-assisted discovery and exploitation had made the old timetable "clearly insufficient". Different words, same judgement: 90 days was sized for human fixing speed, fixing speed has changed, so the timetable has to change.

One background figure is worth adding: the number of published software vulnerabilities in 2026 passed 100,000 on 13 November, double the whole of 2025's total. That figure comes from vulnerability statistics, not from Anthropic or Google, but it shows the scale has changed: if disclosure keeps running at the old pace, the backlog can only grow. Anthropic's own dashboard numbers tell the same story: as of 2 November, its models had found 37,966 candidate vulnerabilities, of which 6,376 had been through human review and 6,009 confirmed as real, a 94.2% true-positive rate. Machines now find vulnerabilities far faster than humans can triage and fix them; the disclosure window is only the last step in that pipeline.

But the other side of the coin matters too. Faster fixing does not mean every project fixes fast. A volunteer-maintained open-source project and a company with a dedicated security team need very different amounts of time for the same bug. The new rules put more pressure on the former — which is exactly what critics are most worried about, as the next section discusses.

Who feels it most: open-source maintainers get 30 days

Under the new rules, the most pressure falls on open-source maintainers — above all the small projects kept alive by volunteers. Fixing a critical vulnerability in 30 days is a scheduling problem for a large company; for a project maintained by one person in their spare time, it can be a question of "whether I have a free weekend this month". Anthropic acknowledges this in its blog post, with a dedicated passage: it will keep including actionable guidance with each report, along with a patch or mitigation where possible, and will agree on a sustainable pace with a maintainer before sending a large volume of findings to a single project. Those promises are written in the blog; whether they hold up will be judged by how 2027 actually plays out.

Anthropic also lists three programmes meant to help maintainers keep up with the new pace. The first is OSS Scanner, a no-cost service that periodically scans critical open-source projects with its strongest models and sends findings directly to maintainers — in effect, scanning your code before someone outside finds the bug. The second is the Claude for Open Source Program, under which eligible maintainers and contributors can apply for six months of Claude Ultra at no cost. The third is the Cyber Verification Program, which tailors Claude's safeguards to enable legitimate defensive work by security professionals, including open-source maintainers. Together, the three programmes are Anthropic's subtext: we cut the time, but we also handed over tools.

One warning, though: OSS Scanner is itself a controversial service. When Anthropic introduced it in October, it stated plainly that its reports go directly to maintainers without human review — while this new policy explicitly says it does not apply to AI-discovered findings that have not received human review. In other words, OSS Scanner findings and the "reports" under this policy are two different tracks; do not mix them up. VibeFix already published an article in October about OSS Scanner, about the free scanning service itself; this article is about the change to the disclosure clock. They are two different stories.

Look at the dashboard numbers, and the reality for maintainers is not encouraging: as of 2 November, Anthropic had disclosed 5,087 vulnerabilities across 680 open-source projects, of which only 629 had been patched; of the reports sent out, 3,635 had been reported to maintainers, and 733 had received CVE or GitHub Security Advisory identifiers. 629 against 5,087 says something plain: for most disclosed vulnerabilities, no patch exists yet. Cutting the window to 30 days adds pressure on top of that reality — if maintainers cannot keep up, more vulnerabilities will be public while the share with patches may fall further. That is the core of the argument between supporters and critics, and the dashboard figures for 2027 will settle it.

Exceptions and boundaries: not every vulnerability gets 30 days

The new policy is not one-size-fits-all, and several exceptions are worth remembering. First, some vulnerabilities simply cannot be fixed fast: flaws in hard-to-deploy software such as critical infrastructure, and flaws that require changes to a specification or RFC. Anthropic says in those cases it will work with the vendor on a remediation timeline rather than enforcing 30 days. That is a concession to reality — and also the part of the new policy most likely to cause argument, since the blog does not say who decides what counts as a reasonable timeline.

Second, the new policy does not apply to AI-discovered findings that have not received human review. A vulnerability a model found on its own, which no human has yet looked at, does not start this 30-day clock. That line matters: it separates "AI-assisted discovery" from "AI-automated discovery", and answers the worry that machine-generated reports would flood maintainers. OSS Scanner findings fall in this category and are not covered by the new policy.

Third, the time boundary: the new timelines apply only to vulnerabilities reported on or after 1 January 2027. Reports already in flight keep their existing deadlines; the clock does not suddenly speed up because the policy changed. That is good news for maintainers working through reports right now: your current countdown will not shrink overnight. But equally, any vulnerability reported after 1 January 2027 starts on a 30-day clock, with no transition period.

Fourth, the backlog: as noted above, the overdue high- and critical-severity reports get an additional 30 days' notice, then are published from 1 January 2027. Maintainers in that group should act now — check the Anthropic CVD dashboard for your project and confirm the status of the report, rather than finding out on the day it goes public.

One last detail: Anthropic waits 30 days after a patch before publishing full technical details — so even after a vulnerability is fixed, the details are not public immediately. Those 30 days are for users to install the patch. Project Zero's new policy works differently: once the day count expires, the details go public whether or not a fix exists. Both end up at 30 days, but Anthropic's 30 days runs from the report and still leaves 30 days after the patch, while Project Zero's 30 days means "public when the clock runs out, fixed or not". Same number, different temper: Anthropic's version is friendlier to users, Project Zero's tougher on vendors.

The other side: three worries from the critics

The new policy had barely landed before the criticism arrived, and three worries deserve to be taken seriously. The first is patch quality. Will a 30-day deadline for a critical vulnerability push teams to ship a patch that has not been properly tested? Patches that introduce new vulnerabilities are not rare. Anthropic's answer is that its reports will include a patch or mitigation to help maintainers cut corners safely — but "including a patch" is not the same as "the patch is right", and the responsibility for testing and shipping still sits with the maintainer. The tighter the deadline, the easier it is for testing to get squeezed. That is a real trade-off, not something a slogan about "security first" can wish away.

The second worry is that small projects will be publicly exposed. A library maintained by one or two volunteers may not be able to fix a critical vulnerability in 30 days — and then the details go public while no patch exists. The published details can be used by defenders to write detection rules, but also by attackers to write exploits. Whoever moves first wins. For a small project, the day of publication is the most fragile moment. Anthropic says it will agree on a "sustainable pace" with maintainers and talk before sending large volumes of reports, but those promises are not written into hard rules, and how much they bind in practice will only be visible in real cases during 2027.

The third worry is the most fundamental: does shortening the disclosure window help defenders, or attackers? Supporters argue the vulnerability would be rediscovered anyway, so earlier disclosure means earlier defence. Critics argue that after disclosure, the users without a patch are the most exposed — and patch installation has never kept pace with publication. Project Zero's reason for dropping the 30-day soak time is that "once a patch is out, attackers can work out the bug from it", so the buffer is meaningless; critics reply that the buffer was never about whether attackers know about the bug, but whether users have had time to install the patch. The disagreement is, at bottom, a different judgement about which side time favours. There is no standard answer.

There is also an angle that is often overlooked: these two organisations are both the rule-makers and the finders. Anthropic uses its own models to find vulnerabilities and sets its own disclosure clock; Google's Project Zero does the same. When the finder sets the rules, it is fair to ask whether the timetable is set to the finder's rhythm rather than the fixer's. Anthropic's blog spends considerable space explaining "why 30 days", but the explanation always starts from "discovery got faster", never from "fixing got faster". Whether fixing has actually kept up with discovery is an open question — the dashboard's 629 patched against 5,087 disclosed suggests it has not, at least not yet.

None of this means the new policy is wrong. A 90-day window was designed for an older speed of discovery, and clinging to it would make the quiet period look increasingly like self-comfort. But the risks the critics point to are real, too: after the rules change, the first to feel the pressure will not be large companies, but the volunteer-maintained open-source projects that have no dedicated security team. If 2027 produces a case where "the vulnerability went public, no patch existed, and the attack arrived first", the argument will be right back where it started. The policy text is settled; the real answer is in next year's dashboard numbers.

What it means for developers: three things you can do now

First, check whether projects you maintain or depend on appear on Anthropic's dashboard. It lists 5,087 disclosed vulnerabilities and 733 CVE and GHSA identifiers across 680 projects. If you maintain an open-source project, check whether it is there and whether any reports are pending; if you are a user, check whether libraries you depend on have disclosed-but-unfixed vulnerabilities. The dashboard is at red.anthropic.com/2026/cvd/, with data as of 2 November and updates to follow.

Second, reassess your patching rhythm. If your team currently takes more than 45 days on average to handle a critical vulnerability, then from 1 January 2027, vulnerabilities reported by Anthropic will be public after 30 days — fixed or not. That means patching, testing and release all have to get faster, or at minimum you need a plan for "mitigation within 30 days, proper patch to follow". A mitigation — switching off a feature, adding a filter rule — is not the same as fixing the bug, but against a publication countdown it buys time.

Third, consider the free tools Anthropic offers. OSS Scanner is free; the Claude for Open Source Program offers six months of Claude Ultra at no cost. These are concrete help. OSS Scanner in particular turns "waiting passively for reports" into "finding things first yourself" — it scans your code before someone outside finds the bug. Before using it, think it through: its reports go to you directly without human review, so you will have to deal with false positives yourself.

One final judgement. Is 30 days enough to fix a critical vulnerability? For a large company, yes; for a small project, tight. But the direction of the change is right: the speed of discovery has already changed, and if the disclosure system does not move with it, the quiet period only buys time for attackers. The real test is not in the policy text but in how 2027 is executed: whether Anthropic delivers on its promise to include patches and mitigations, whether maintainers can keep up with the 30-day rhythm, and whether that dashboard figure — 629 patched — starts climbing. These numbers are worth coming back to check in three months.

Views 1Comments 0

Comments (0)

ME
0/1000
Loading comments...

Sources

Browse projectsPublish your project

Related articles

An analyst seen from behind in a dark security operations center, facing large screens showing world maps and network topology
News
The Patch Just Shipped and the Exploit Is Already There: Google Finds the First Fully Autonomous AI Exploitation Framework

On 16 November 2026, Google's Threat Intelligence Group reported the first fully automated vulnerability discovery and exploitation chain: no human in the loop. The evidence is an exposed server running exploit_agent, a framework that diffs code around patch announcements, writes exploits and fires them. The same report shows a North Korean cluster turning Claude Code's settings file into a backdoor trigger — and stolen cloud compute running up $100,000 bills in a day.

Industry TrendsAI CodingSecurity & Privacy
A modern data center corridor lined with black server racks, cable trays and ventilation ducts overhead, blue indicator lights blinking on the racks
News
Two Months Late, PostgreSQL 19 Is Finally Here: Repack Tables While They Run, and Make Standbys Wait for Your Writes

On 19 November 2026, PostgreSQL 19 shipped — two months late, after a flood of AI-reported vulnerabilities rewrote the schedule. It brings REPACK (CONCURRENTLY), pg_plan_advice, and WAIT FOR LSN: tables can be reorganized while they run, query plans can be pinned, and standbys wait for your writes.

Product LaunchData WorkflowsBackend Engineering