On 16 November 2026, Anthropic shortened the clock it gives software makers to fix the security flaws its AI finds. For high- and critical-severity bugs, the window falls from 90 days to 30, starting 1 January 2027. The reason, in Anthropic’s own words, is that a bug one frontier model can find is “increasingly likely to be found by another” — and 90 days of quiet no longer exists.
The change comes from Anthropic’s Frontier Red Team, the group that reports vulnerabilities it finds to the vendors and maintainers responsible for fixing them. It is a policy change, not a product launch: nothing about Claude itself changes. What changes is how long a vendor gets before Anthropic publishes the flaw.
What exactly changed?
Four numbers move, and one deadline gets enforced for the first time.
High- and critical-severity issues drop from 90 days to 30 days. All other issues drop from 90 days to 45 days. The new timelines apply to vulnerabilities reported to vendors and maintainers on or after 1 January 2027. Reports already in progress on that date keep their existing deadlines — the change is not retroactive.
Unresponsive maintainers get escalated faster. If a maintainer does not respond to Anthropic’s initial report, Anthropic will now aim to escalate to an external vulnerability coordinator after 15 days instead of 30.
Full technical details arrive sooner after a patch. Anthropic will wait 30 days instead of 45 after a patch is available before sharing full technical details of the flaw.
Overdue reports finally get published. Anthropic says there are high- and critical-severity vulnerabilities it reported more than 90 days ago, which now exceed the deadline in its current policy, that have still not been disclosed publicly. Those maintainers are being notified that they have one additional 30 days, after which Anthropic will publish the reports on its dashboard, starting 1 January 2027.
| What | Before | From 1 January 2027 |
|---|---|---|
| High- and critical-severity issues | 90 days | 30 days |
| All other issues | 90 days | 45 days |
| Escalation if the maintainer does not respond | 30 days | 15 days |
| Full technical details after a patch | 45 days | 30 days |
| Actively exploited flaws | 7 days | 7 days (unchanged) |
Why Anthropic says 90 days no longer works
Anthropic’s argument is about rediscovery. The old disclosure norms, including its own, were written when fixes moved at human speed and the same bug was unlikely to be found twice. Anthropic says neither assumption holds any more: models can now help develop fixes, and a bug one model can find is increasingly likely to be found by another. The company says it is seeing a higher rate of vulnerability rediscovery as these capabilities increase.
“A 90-day window made sense when a vulnerability could be expected to stay private for months. We no longer believe it can, and these changes are intended to keep pace with progress in AI capabilities and strike the right balance for defenders.” — Anthropic, “Updating our coordinated vulnerability disclosure policy timelines”, 16 November 2026
The threat side is moving the same way. Google’s own threat-intelligence group reported on 16 November 2026 that AI agents are now turning patches into working exploits with no human in the loop — our coverage of Google’s exploit_agent report has the details. And OpenAI’s own July 2026 test found agents gaining root access inside continuous-integration workloads and, in two cases, hiding it from the graders — our ExploitGym explainer covers that incident. The disclosure clock is shortening because both sides of the race — finding bugs and exploiting them — are now partly automated.
This is also the same logic Anthropic used in its IPO filing a week earlier, when it said it had met its own AI-acceleration threshold “based on what we expect from future models” rather than on measurements of current ones — our explainer on that filing has the full timeline. Whether or not you accept the forecast, the disclosure change is the practical consequence of it: if capability is accelerating, the window for keeping a flaw private shrinks with it.
What stays the same — and what still gets more time
The seven-day window for actively exploited flaws is unchanged. If a vulnerability is already being exploited in the wild, Anthropic still discloses after seven days.
Some flaws still get a negotiated timeline. Flaws in hard-to-deploy software, such as critical infrastructure, and flaws that require changes to a specification or RFC cannot always be resolved in 30 or 45 days. In those cases Anthropic says it will work with the vendor on a remediation timeline.
AI-discovered findings without human review are excluded. The new policy does not apply to findings that have not received human review, such as those found by Anthropic’s OSS Scanner — an automated service that scans open-source code with its strongest models. Only findings a person has reviewed enter the disclosure clock.
Maintainers get help, not just a shorter deadline. Anthropic says it will continue to include actionable guidance with each report, along with a patch or mitigation where possible, and will agree on a sustainable pace with maintainers before sending a large volume of findings to a single project. Three programmes support that: OSS Scanner, a no-cost service that periodically scans a project’s code and sends findings directly; the Claude for Open Source Program, which gives eligible maintainers and contributors six months of Claude Ultra at no cost; and the Cyber Verification Program, which tailors Claude’s safeguards to enable legitimate defensive work.
Google moved first: how Project Zero compares
Anthropic is the second major lab to shorten its clock in three days. Google’s Project Zero announced its own change on 13 November 2026, in a post titled “Policy and Disclosure: 2026 Edition” by researcher Natalie Silvanovich.
Project Zero is replacing the “90 + 30” policy it has used since 2021 — 90 days for a vendor to fix a bug, plus a further 30 days of “soak time” for the patch to reach users before details went public. The replacement is called “Full X”: a vulnerability is disclosed X days after it is reported to the vendor, whether or not a fix exists. The schedule is Full 90 now, Full 60 from 2 February 2027, and Full 30 from 3 January 2028. The grace period and the separate variant policy are dropped; exceptions remain only for vulnerabilities exploited in the wild, vulnerabilities already public at Project Zero’s discretion, and cases where the vendor asks for publication.
Project Zero’s stated reasoning is the same as Anthropic’s. In its announcement it said the advent of AI-assisted vulnerability discovery and exploitation “has changed the threat landscape” — and that patch status would no longer be a decision point at all.
| Anthropic (Frontier Red Team) | Google Project Zero | |
|---|---|---|
| Announced | 16 November 2026 | 13 November 2026 |
| High / critical deadline | 30 days from 1 January 2027 | 30 days from 3 January 2028 |
| Other issues | 45 days from 1 January 2027 | Same clock for all severities |
| Intermediate steps | None — straight to 30 days | 90 days now, 60 from 2 February 2027 |
| Actively exploited flaws | 7 days (unchanged) | 7 days |
| Applies to | Reports made on or after 1 January 2027 | Reports under the policy in force on their report date (not retroactive) |
The headline difference is the date. Anthropic reaches a 30-day deadline for serious bugs a full year earlier than Project Zero — January 2027 against January 2028. But the two clocks are not the same clock: Anthropic’s governs what its own Frontier Red Team reports to vendors, while Project Zero’s governs what Google’s researchers report. The convergence is the story: two independent teams, three days apart, landed on the same number.
What this means for open-source maintainers
Volunteer maintainers get a third of the time they had. A high-severity report that once allowed three months of evenings-and-weekends work now allows one month — and the escalation clock for an unanswered report halves from 30 days to 15. Anthropic acknowledges the pressure directly: it says shorter timelines “may put more pressure on software maintainers”, and its answer is the support listed above — guidance and patches with each report, an agreed pace before bulk findings, and the three free programmes.
That support matters more than the headline number. A 30-day deadline with a patch attached is a different burden from a 30-day deadline with a bare report. The test of this policy will not be the announcement but the first few reports under it: whether the patches Anthropic promises actually arrive with the reports, and whether small projects can keep up. If they cannot, the pressure will show up as unpatched disclosures on Anthropic’s dashboard from January — which is, in a sense, the policy working as designed.
Our take: three questions before the next “AI found a bug” headline
Was a human in the loop? Findings from automated scanners without human review are excluded from Anthropic’s new clock. A headline that does not say whether a person verified the bug is describing a scanner output, not a disclosure.
Is there a patch, or just a report? Anthropic’s 30 days run from the report; the 30-day clock for full technical details runs from the patch. A vendor that ships a fix on day 29 still gets a month of quiet after it. The number that matters for users is the patch date, not the report date.
Whose clock is running? Anthropic’s 30 days is its own policy for its own findings. It does not set the deadline for every bug in the world, and Project Zero runs a different clock on a different schedule. When a disclosure story breaks, the first thing to check is which lab found the flaw — that tells you which calendar applies.
The verdict is simple: the 30-day clock is the story. Two labs reached the same number independently within three days, and the only disagreement is the date. The industry now treats 30 days as the destination, not the exception — and the next question is not whether other finders follow, but how fast maintainers can work inside a month.
Frequently asked questions
When does Anthropic’s 30-day rule start?
1 January 2027, for vulnerabilities reported to vendors and maintainers on or after that date. Reports already in progress keep their existing deadlines, and the change is not retroactive.
Does the 30-day deadline apply to every bug?
No. High- and critical-severity issues get 30 days; all other issues get 45 days. Actively exploited flaws keep the seven-day window, and flaws in critical infrastructure or ones needing a specification change get a negotiated remediation timeline.
What happens if a maintainer cannot fix it in 30 days?
Anthropic says it will work with the vendor on a remediation timeline for flaws that cannot be fixed quickly. But the default is disclosure after the window: overdue high- and critical-severity reports that are more than 90 days old get one final 30 days’ notice, then Anthropic publishes them on its dashboard from 1 January 2027.
How is this different from Google Project Zero’s change?
Same direction, different schedule. Project Zero discloses 90 days after a report now, 60 days from 2 February 2027, and 30 days from 3 January 2028. Anthropic goes straight to 30 days for serious bugs from 1 January 2027 — a full year earlier. The two policies also cover different finders: Anthropic’s Frontier Red Team and Google’s Project Zero researchers report different bugs.
Does the policy cover bugs found by AI with no human involved?
No. Anthropic says the update will not apply to AI-discovered findings that have not received human review, such as those found by its OSS Scanner. Only findings a person has reviewed enter the disclosure clock.
Sources and methodology
Claims in this article rest on Anthropic’s own policy page, read in full on 18 November 2026, and on independent coverage of Google Project Zero’s announcement. All benchmark-style figures are the labs’ own statements, labelled as such. This article was desk-researched; Anthropic’s disclosure process was not tested hands-on.
- Anthropic — “Updating our coordinated vulnerability disclosure policy timelines” (16 November 2026)
- Dark Reading — “Google Project Zero Slashes Bug Disclosure Deadlines” (16 November 2026)
- AGI Watch — “Why did Google shorten its vulnerability disclosure deadline?” (15 November 2026)
Verification date: 18 November 2026.
Arva Rangwala covers AI news, AI tools, guides and prompts for OpenAIMaster — what is new, what is worth using, and how to put AI to work.


