Updating our coordinated vulnerability disclosure policy timelines

A Frontier Red Team policy post announcing that, for vulnerabilities reported on or after 1 January 2027, Anthropic will disclose findings its models discover after 30 days for high and critical severity and 45 days for others instead of 90, escalate silent maintainers after 15 days instead of 30, and publish full details 30 days after a patch instead of 45, citing faster AI-assisted fixes and more frequent rediscovery; overdue unpublished reports get 30 more days.

Anthropic (Frontier Red Team) · 2026-11-16 · original

How long should a finder keep a vulnerability private once a model, not a person, found it? This short post, published on 16 November 2026 by Anthropic's Frontier Red Team with no individual byline, announces shorter deadlines in Anthropic's coordinated vulnerability disclosure (CVD) policy, the rules it follows for vulnerabilities its models find in other people's software. The post is about 600 words and reports no measurements. The version of record has two parts: the post, and the policy page itself (anthropic.com/coordinated-vulnerability-disclosure), which carried "Last updated Mar 6, 2026" in Wayback Machine captures of 9, 16, 19 and 20 November 2026 (the last at 12:05 UTC) and "Last updated Nov 20, 2026," with the new terms, when read on 21 November 2026.

What changes

The post lists four changes (opening section); the policy page's revision matches each:

Term Before After
Public disclosure deadline, high and critical severity 90 days 30 days
Public disclosure deadline, all other issues 90 days 45 days
Escalation to an external vulnerability coordinator when a maintainer does not respond after 30 days after 15 days
Wait after a patch is available before full technical details 45 days 30 days

The new timelines apply to vulnerabilities reported on or after 1 January 2027; reports in progress on that date keep their deadlines. For high- and critical-severity reports already past the old 90-day deadline and not yet public, Anthropic is giving maintainers 30 more days, "after which time we will publish these reports on our dashboard in line with our updated policy starting January 1, 2027." The post does not say how many reports are in that group.

Several terms are unchanged, according to the post: the seven-day window for vulnerabilities under active exploitation (whose scope the policy page's wording widened; see "The written policy text"), the possibility of agreeing a longer remediation timeline for "hard-to-deploy software, such as critical infrastructure," and for flaws needing changes to "a specification or RFC," and the practice of agreeing a pace with maintainers before sending many findings to one project. The policy page keeps its 14-day extension on request for maintainers making progress.

The reasons given

The post argues that disclosure norms, including Anthropic's own, "assumed that fixes moved at human speed and that independent rediscovery of the same bug was unlikely," and that "Neither assumption holds today" ("Responding to AI-enabled vulnerability discovery"). Its grounds are that models "can now help develop fixes," that "a bug that one model can find is increasingly likely to be found by another," and that Anthropic is "seeing a higher rate of vulnerability rediscovery." It gives no rediscovery rate, no time-to-fix figures and no count of affected reports. The conclusion is stated as a belief: a 90-day window assumed a vulnerability "could be expected to stay private for months. We no longer believe it can."

The post acknowledges that "shorter timelines may put more pressure on software maintainers" and lists what Anthropic offers in return: actionable guidance and, where possible, a patch or mitigation with each report; OSS Scanner, its free scanning service for critical open-source projects; six months of Claude Ultra at no cost through the Claude for Open Source Program; and its Cyber Verification Program for security professionals. These are Anthropic's own products and programs.

Scope

The policy covers vulnerabilities Anthropic discovers in open-source software, and in closed-source software where it has authorization to do security research (policy page, "Scope"). The post states that the update "will not apply to AI-discovered findings that have not received human review, such as those found by OSS Scanner." Those reports go to maintainers unvalidated and, under the OSS Scanner FAQ, carry no disclosure deadline until Anthropic validates them (see An opt-in vulnerability-finding service for open-source software).

The written policy text

The policy page is the governing document; the post summarizes it. A word-level comparison of the 19 November capture with the page of 21 November shows the four changes above, a new "Effective Date" paragraph, and three further edits the post does not list:

  • The "General" statement changed from "aims to follow the industry standard 90-day disclosure deadline" to "aims to follow industry vulnerability disclosure norms."
  • The active-exploitation bullet changed from "For actively exploited critical vulnerabilities, we target a 7-day timeline" to "For any actively exploited vulnerabilities, we target a 7-day disclosure timeline."
  • "90-day deadline" in the extension bullet became "disclosure deadline."

The second edit bears on a sentence in the post. The post says that "Our seven-day disclosure window for any vulnerability under active exploitation remains unchanged" ("Responding to AI-enabled vulnerability discovery"), but the March text applied the 7-day target to "actively exploited critical vulnerabilities," and the revised page extends it to "any." The March text did already say, in its bullet on severity disagreements, that with "credible evidence of active exploitation" the 7-day target "would apply regardless of how the maintainer classifies the bug"; that bullet concerns the maintainer's classification, and the narrower wording stood in the main bullet. The length of the window is unchanged; its stated scope widened.

The disclosure sentence reads, in the new text, that Anthropic aims "to share details publicly with defenders after 30 days for Critical and High vulnerabilities, and 45 days for all other vulnerabilities, or after a patch is released, whichever comes first," while the separate "Patch Details" paragraph now sets a 30-day wait after a patch before "full technical details." The page does not define how "details" differ from "full technical details"; the March text had the same two-part structure.

Limitations

  • No evidence in the post. The rediscovery and fix-speed claims are asserted, not measured; the CVD dashboard the post links had last been updated on 2 November 2026 when read for this note, and its overview reports counts of candidate, disclosed and patched findings but no rate of rediscovery by others.
  • Interest. Anthropic is both the finder whose deadlines shorten and the provider of the scanning and model programs offered to maintainers.
  • Unknown size. The number of overdue high- and critical-severity reports that will be published is not given.

Research questions

  1. How often are Anthropic's findings rediscovered, and how fast? Publishing the share of reported vulnerabilities independently reported by others before disclosure, by quarter, would test the post's central premise.
  2. Do shorter deadlines change time to patch? Comparing time from report to fix for reports made before and after 1 January 2027, by severity and project size, would show whether maintainers keep pace.
  3. How many reports are published under the 30-day grace period, and how many were patched by then? The dashboard's ledger could answer this after early 2027.
  4. How often is the remediation-timeline exception used? Counts of extensions granted for critical infrastructure and specification-level flaws would show how much the default deadlines bind.

How it connects

References