Anthropic will shorten coordinated vulnerability disclosure (CVD) windows to 30 days for high- and critical-severity flaws and 45 days for everything else, because frontier models are finding and re-finding bugs faster than old 90-day norms assumed. The November 16, 2026 Frontier Red Team post sets January 1, 2027 as the start date for new reports while leaving in-flight cases on their existing clocks.
The change lands the same day Google GTIG documented exploit_agent, an autonomous patch-to-exploit pipeline in the wild. Anthropic is answering the defender side of that pressure: if machines rediscover flaws quickly, long private embargoes may not protect users.
TL;DR
| Topic | New rule (from Jan 1, 2027) | Old rule |
|---|---|---|
| High / critical disclosure | 30 days | 90 days |
| Other severities | 45 days | 90 days |
| Coordinator escalation | 15 days without maintainer response | 30 days |
| Details after patch ships | 30 days | 45 days |
| Active exploitation | 7 days | 7 days (unchanged) |
Agents that act on your behalf need rules. See how explainx helps secure your agents →
What Anthropic actually announced
Anthropic's post is short and operational. It does not publish a diff PDF; it states policy deltas in plain language.
Shorter embargo windows. High- and critical-severity issues move from a 90-day coordinated window to 30 days. All other severities move to 45 days. Reports filed on or after January 1, 2027 use the new schedule.
Faster third-party escalation. If a maintainer does not answer an initial report, Anthropic will try to involve an external coordinator at 15 days instead of 30.
Earlier public technical detail. After a fix is available, Anthropic will release full write-ups 30 days later instead of 45.
Backlog cleanup. Anthropic says some high- and critical issues reported more than 90 days ago never reached public disclosure under the old policy. It is notifying those maintainers of an extra 30 days, after which findings may appear on Anthropic's dashboard under the updated rules.
Human review gate for OSS Scanner. AI-only hits from OSS Scanner without human review do not fall under this update — a line that matters for open-source maintainers flooded with automated reports.
Why Anthropic says 90-day CVD no longer fits
A gear with a green bandage over a crack, evoking faster security patch and disclosure cycles in the AI era
Anthropic's reasoning rests on three shifts:
- Models help find and fix bugs — which Anthropic frames as good for defenders.
- Models also rediscover the same bugs — so private knowledge does not stay private long.
- Fix speed assumptions changed — policies written for human-only research assumed months of safety; that assumption "no longer holds," in Anthropic's words.
The post still acknowledges slow infrastructure: flaws in hard-to-deploy systems or those needing specification changes may need custom timelines. Anthropic says it will negotiate with vendors in those cases.
That nuance matters for maintainers of embedded firmware, telecom gear, or standards-track protocols. The default clock is faster; the exceptions are negotiated, not automatic extensions.
What changes for open-source maintainers
A figure with a green lantern guides many small shapes, evoking maintainer duty under faster disclosure clocks
Anthropic explicitly worries about pressure on volunteers. Its mitigations are practical, not rhetorical:
- Actionable reports with patches or mitigations where possible.
- Pacing agreements before dumping large finding batches on one project.
Programs named in the post give maintainers tooling to keep up:
- OSS Scanner — periodic scans with Anthropic's strongest models, no cost for eligible critical projects.
- Claude for Open Source Program — six months of Claude Ultra for eligible contributors.
- Cyber Verification Program — adjusted safeguards for legitimate defensive work.
If you maintain a widely depended-on library, the combined message is: expect more AI-generated reports and shorter clocks if Anthropic or a peer finds the same bug.
Timeline comparison for defenders and maintainers
Most teams still plan patch and disclosure cycles around quarterly releases. Anthropic's calendar is now closer to weekly escalation for critical issues once a vendor misses email.
| Day | Maintainer action | Anthropic action (new policy) |
|---|---|---|
| 0 | Receive report with repro and optional patch | Ships actionable guidance when available |
| 7 | Patch or mitigation plan for active exploitation | Public disclosure if exploitation is live |
| 15 | Respond to initial thread | Escalate to external coordinator if silent |
| 30 | Ship fix for high or critical issues | Disclose high or critical if still unfixed |
| 45 | Ship fix for lower severities | Disclose other severities if still unfixed |
| 30 after patch | — | Publish full technical detail (was 45 days) |
This table is explainx.ai's reading of the Frontier Red Team post. Your counsel may map severities differently than Anthropic's internal ratings.
Enterprise buyers should ask vendors new questions
If you procure foundation models, agent platforms, or security scanners, add these questions to RFPs before January 2027:
- What CVD window do you commit to when a frontier model reports a flaw in your product?
- Do you accept coordinated reports that include machine-generated drafts from OSS Scanner-class tools?
- Who signs off before a partially automated report goes to a maintainer — and is that human named in your SLA?
Anthropic is publishing its own answers. Your CMS, EHR, or IAM vendor may still cite 90-day embargoes written years ago.
How this fits the wider disclosure debate
A rounded envelope with a small green padlock, suggesting coordinated vulnerability disclosure under shorter timelines
explainx.ai has covered adjacent threads:
- Google GTIG on shrinking n-day windows when attackers automate patch diffs.
- Vulncheck and Glasswing on fast patching versus KEV metrics.
- Claude Security scan-to-fix for enterprises.
Anthropic's move nudges the ecosystem toward defender AI racing attacker AI on calendar time, not just on benchmark scores.
Regulators and customers may soon ask vendors: What is your CVD SLA when the finder uses a frontier model? Anthropic just published its answer.
How Anthropic CVD compares to Google and Microsoft norms
Google's Project Zero popularized 90-day disclosure for many issues, with shorter clocks for actively exploited bugs. Microsoft's Coordinated Vulnerability Disclosure philosophy similarly balances vendor prep time with user risk.
Anthropic's 30/45-day default, starting in 2027, is shorter and explicitly ties the change to model rediscovery, not just human researchers. That is a new variable: if two labs' models can find the same flaw independently, the private embargo protects fewer people for fewer days.
Nothing stops a maintainer from negotiating extensions for infrastructure that cannot patch in a month. Anthropic says it will work on custom timelines for hard-to-deploy systems. Document those negotiations in writing so your security counsel can cite them if public disclosure nears.
Rediscovery risk in plain language
Imagine Lab A's model reports a critical RCE to a vendor on January 5, 2027. The vendor is silent. On February 4, Anthropic may disclose under the new high-severity clock.
If Lab B's model — or an attacker’s exploit_agent-style pipeline — finds the same bug on January 20, users were exposed for two weeks before either public or private knowledge converged. Shorter CVD does not fix rediscovery; it shrinks how long silence is allowed.
That is why pairing Anthropic's policy with faster patching and runtime agent controls matters more than debating whether 30 or 45 days is fair.
What security teams should do before January 2027
- Rebaseline SLAs with upstream vendors; 90-day assumptions in runbooks are stale.
- Prepare comms templates for partial disclosures at day 30 when fixes slip.
- Split workflows for human-validated versus machine-generated reports to avoid alert fatigue.
- Track Anthropic's public dashboard after the grace period for overdue critical issues.
- Pair faster disclosure with faster patch deployment — shorter embargoes help only if your CI/CD can ship fixes in days.
Open-source maintainer playbook under shorter clocks
Volunteer maintainers bear the brunt of shorter windows. Anthropic's mitigations — paced delivery, patches bundled with reports, OSS Scanner for eligible repos — only help if you opt in and document your response process.
Day 0–3: Acknowledge receipt, even if the fix needs research. Silence triggers coordinator escalation at day 15.
Day 3–10: Reproduce on the minimum supported release branch. If the report includes a patch, test it against your CI matrix before merging.
Day 10–25: Ship a release candidate or mitigation (config flag, disable feature) for high or critical issues. Communicate ETA publicly if the full fix needs a major bump.
After patch: Expect full technical detail from Anthropic 30 days later instead of 45 — plan your own advisory blog accordingly.
If you run a critical dependency, apply to OSS Scanner before January so automated findings arrive with human review already applied where Anthropic promises it.
What would change our read
Three signals would show the policy is enforced, not aspirational:
- Public dashboard entries on day 30 for overdue critical issues named in Anthropic's post.
- Maintainer feedback at scale — painful, but evidence the clock is real.
- Matching disclosure calendars from OpenAI, Google, or Microsoft, turning one lab's rule into industry practice.
A delay past January 1, 2027 or broad new exemptions for AI-only reports would weaken the signal.
