Anthropic cuts bug disclosure window to 30 days for high and critical flaws
From 1 January 2027, Anthropic will publish high and critical vulnerabilities its Claude models find in open-source software after 30 days instead of 90, citing faster fixes and bug rediscovery.
· 7 min read
Written by AI from the sources linked below. How we use AI
On this page
Anthropic said on 16 November 2026 that it is cutting the public disclosure window for vulnerabilities its Claude models find in open-source software from 90 days to 30 days for high and critical severity issues, and to 45 days for everything else. The new rules apply to reports sent to maintainers on or after 1 January 2027. For maintainers who already receive AI-generated bug reports, the change shortens the time they have to ship a fix before details become public.
Key points
- For high and critical severity findings, Anthropic's disclosure deadline drops from 90 days to 30 days; all other findings get 45 days.
- The new timelines apply to vulnerabilities reported to vendors and maintainers on or after 1 January 2027; reports already in progress keep their current deadlines.
- Escalation to an external coordinator after a silent maintainer moves from 30 days to 15 days, and full technical details follow a patch after 30 days instead of 45.
- Anthropic's seven-day window for actively exploited vulnerabilities is unchanged, as is the rest of the policy.
- High and critical reports made more than 90 days ago, past the current deadline and still undisclosed, get 30 more days from notification before publication on Anthropic's dashboard.
What happened
In a post by its Frontier Red Team on 16 November 2026, Anthropic set out the new "coordinated vulnerability disclosure" (CVD) timelines. The company's reasoning is that two assumptions behind the old norm no longer hold. Earlier practice assumed fixes would move at human speed and that independent rediscovery of the same bug was unlikely. The post says "Neither assumption holds today." Anthropic says models can now help write fixes, and that a bug one model finds is increasingly likely to be found by another, so it sees a higher rate of rediscovery. A 90-day window, it says, made sense when a vulnerability could be expected to stay private for months.
The concrete changes in the announcement:
- Disclosure deadline: 90 days becomes 30 days for high and critical severity, and 45 days for all other issues.
- Escalation: if a maintainer does not respond to the first report, Anthropic aims to escalate to an external vulnerability coordinator after 15 days (the published policy page, last updated 6 March 2026, says 30 days).
- After a patch: full technical details are shared 30 days after a patch is available, down from 45.
- Unchanged: the seven-day target for actively exploited vulnerabilities.
- Backlog: high and critical reports more than 90 days old that are past the current deadline and undisclosed get an extra 30 days from notification, then are published on Anthropic's dashboard.
Anthropic names exceptions. Hard-to-deploy software, such as critical infrastructure, and flaws that need a specification or RFC change may not be fixable in 30 or 45 days; in those cases it says it will work with the vendor on a remediation timeline. The policy also does not apply to AI-discovered findings that have had no human review, such as those from its OSS Scanner.
On the maintainer side, Anthropic acknowledges that shorter timelines add pressure. It says each report will include actionable guidance and, where possible, a patch or mitigation, and that it will agree a sustainable pace with a maintainer before sending a large volume of findings to one project. It points to three support programmes: OSS Scanner, a no-cost service for critical open-source maintainers; a Claude for Open Source programme offering six months of "Claude Ultra" to eligible maintainers and contributors; and a Cyber Verification Program that tailors Claude's safeguards for legitimate defensive work.
Background
Disclosure windows of around 90 days became a common industry norm after Google Project Zero popularised the approach. Anthropic's published policy, last updated on 6 March 2026, says details are shared "after 90 days, or after a patch is released, whichever comes first," with a 14-day extension available on request when a maintainer is engaged and making progress. Full technical details generally wait 45 days after a patch so downstream users can deploy it. For actively exploited critical vulnerabilities, the policy sets a seven-day target, with a further seven days if the maintainer is actively working on a fix and asks for more time. It also says reports generally reflect findings that a human security researcher has reviewed and confirmed.
The volume behind the policy is large. A VentureBeat report dated 9 October 2026, using data as of 2 October, said Anthropic had reported 6,157 findings across 591 open-source projects, with 516 patched upstream (about 8.4%) and 29,439 findings discovered between 1 November 2025 and 2 October 2026. It also reported that reports had grown 3.9 times since 22 May while patches grew 5.3 times. In that coverage, Todd Ouska of wolfSSL said of the reports his project received: "all but two were valid, and five became CVEs."
Not everyone found the earlier policy clear. In an analysis published on 9 June 2026, VulnCheck's Patrick Garrity argued that the policy mixed timelines (a 90-day publication window, a 45-day wait after a patch, 14-day extensions and "extenuating circumstances" adjustments) and that the practical clock was hard to follow. The same analysis counted 10 findings past the 90-day window at that date, with 168 more due to reach it within about 30 days. A shorter, simpler schedule responds to part of that criticism, although the new post still allows exceptions.
The numbers
Anthropic's public dashboard, as of 2 November 2026 at 19:30 UTC, shows 5,087 vulnerabilities disclosed across 680 open-source projects. Of those, 3,635 have been acknowledged by a maintainer and, to the best of Anthropic's knowledge, 629 have been patched. The dashboard lists 733 identifiers issued: 230 CVEs and 503 GitHub Security Advisories (one finding can have both).
Read the chart as a snapshot of the gap that motivates the change: far more findings have been disclosed or acknowledged than patched. The dashboard's counts and the October VentureBeat figures (6,157 reported, 591 projects) use different labels and dates, and we have not tried to reconcile them. The dashboard also reports that outside security firms reviewed 6,376 findings and confirmed 6,009 as valid (94.2%), and it gives 37,966 candidates found in total. It does not state a total of high and critical findings among those disclosed, nor how many are past deadline, so we cannot say how many reports the shorter 30-day clock would catch.
What it means
For maintainers of widely used projects: a high or critical report from Anthropic will, from 2027, come with a 30-day clock rather than 90. Anthropic says each report will carry guidance and, where possible, a patch. Projects with thin staffing may still struggle, which is the pressure the company concedes. The 15-day escalation also means a project that does not reply risks the finding going to a coordinator sooner.
For users and downstream vendors: a shorter window means details can become public sooner, including before a distribution or vendor has shipped an update. The 30-day wait after a patch is the time left to deploy. Teams that track open-source dependencies should expect advisories from this source to arrive faster.
For the wider industry: Anthropic ties the change to model capabilities rather than to a specific incident. Whether other AI labs and security teams follow, and whether Google's and other vendors' windows shift, is something we could not find stated in sources we could verify.
What we don't know yet
Anthropic's post leaves several points open, and we did not find answers elsewhere:
- Whether the 15-day escalation applies to all severity levels or only high and critical.
- How "high" and "critical" are defined for this policy, and which external coordinator is used.
- Exactly when the 30-day backlog extension clock starts and ends.
- How the seven-day exploitation window interacts with the new windows.
- Eligibility rules for the Claude for Open Source and Cyber Verification programmes, and what "Claude Ultra" includes.
- The full text of the updated policy: the policy page we read (last updated 6 March 2026) still shows the older terms, and the announcement links to the policy rather than reproducing it.
We found no independent coverage of the change at the time of writing, so the terms above come from Anthropic's own post, which is the primary source.
What to watch next
- A revised policy page in place of the 6 March 2026 text, and what it says on the open questions above.
- Maintainer reaction to the 30-day clock, especially from projects that have already received large batches of reports.
- The dashboard's severity breakdown and how many high and critical reports reach the extended backlog deadline.
- Whether other organisations that use AI to find vulnerabilities shorten their own windows.
Sources
- Anthropic: coordinated vulnerability disclosure timelines (16 Nov 2026)
- Anthropic: coordinated vulnerability disclosure for Claude-discovered vulnerabilities (policy page, last updated 6 Mar 2026)
- Anthropic: coordinated vulnerability disclosure dashboard
- VentureBeat: Anthropic's cyber mission starts with 6,157 findings reported to maintainers and 516 patched
- VulnCheck: Anthropic ledger analysis
Written by
The official account of the people who run audiotale.net: announcements, conversation starters and the newsroom's drafts. Posts written with AI say so. Every other account here is a real person.
Read next

Cisco Project Quarry open source rules detect AI agent attacks across a session
Cisco's new open source rule language checks whole agent sessions, not single messages, to catch multi-step attacks such as a file read followed by an external send.
audiotale.net team7 min read
CVE-2026-114595: EPICS Base flaws, found by an AI agent in under 4 hours, hit lab control systems
CISA advisory ICSA-26-300-01 lists four flaws in EPICS Base 7.0.10 and earlier, one rated CVSS 9.8; the fix is 7.0.11.
audiotale.net team7 min read
Google GTIG: exploit_agent turns patches into exploits, no human involved
Google's threat team says attackers ran an AI framework that diffs patches, writes exploits and runs them automatically, the first fully autonomous pipeline it has seen in the wild.
audiotale.net team8 min read