Anthropic is shortening coordinated vulnerability disclosure (CVD) deadlines because frontier models find bugs faster and rediscover the same flaws more often—high and critical issues will move from a 90-day private window to 30 days for reports sent on or after January 1, 2027. The November 16, 2026 policy post does not change how Claude Code runs on your laptop today, but it changes how quickly public write-ups and proof-of-concept detail may appear for libraries, CLIs, and MCP servers your agents import.
| Question | Short answer |
|---|---|
| Am I affected? | Indirectly, if you ship or depend on open-source packages Anthropic reports on |
| Is there a fix? | No product patch—adjust patch SLAs and watch disclosure feeds |
| Does data leave my machine? | Not from this policy change alone |
| When does it start? | Reports to vendors on or after 2027-01-01 |
What Anthropic changed
Anthropic lists four concrete policy shifts in its announcement:
- Disclosure window: 30 days for high and critical severity (down from 90); 45 days for all other severities (down from 90).
- Effective date: Applies to vulnerabilities reported to vendors and maintainers on or after January 1, 2027. In-flight reports keep existing deadlines.
- Escalation timing: Aim to involve an external vulnerability coordinator after 15 days without maintainer response, instead of 30.
- Post-patch detail: Publish full technical details 30 days after a patch is available, instead of waiting 45 days.
Anthropic also addresses backlog: high and critical issues reported more than 90 days ago that exceed the old deadline but were not yet public will get an extra 30 days notice to maintainers, then publication on Anthropic's disclosure dashboard in line with the updated policy starting January 1, 2027.
Unchanged pieces matter for defenders:
- Seven-day disclosure for vulnerabilities under active exploitation stays the same.
- The update does not apply to AI-discovered findings without human review, such as OSS Scanner output.
- Anthropic will still negotiate longer timelines for hard-to-deploy fixes (critical infrastructure, specification changes).
Old versus new deadlines (at a glance)
| Topic | Previous policy | From reports on or after 2027-01-01 |
|---|---|---|
| High / critical disclosure | 90 days | 30 days |
| Other severities | 90 days | 45 days |
| Maintainer non-response escalation | ~30 days | ~15 days to external coordinator |
| Full detail after patch ships | 45 days | 30 days |
| Active exploitation | 7 days | 7 days (unchanged) |
| Unreviewed OSS Scanner output | Not covered by this update | Still excluded |
Reports already in flight on January 1, 2027 keep their old deadlines, which matters if you maintain a library where Anthropic filed months ago under the 90-day clock.
Why Anthropic says CVD norms broke
The post argues coordinated disclosure rules assumed human-speed patching and low independent rediscovery risk. Anthropic now states both assumptions fail: models help develop fixes, and bugs one model finds are increasingly found by another. A 90-day quiet period "made sense when a vulnerability could be expected to stay private for months"; Anthropic no longer believes it can.
That aligns with external data Beam has covered:
- CVE volume in 2026 rising alongside AI-assisted discovery.
- Google GTIG's Q4 tracker documenting exploit_agent patch-differential automation in the wild.
- Anthropic's own prior research on AI-built exploit chains in the September Chrome and kernel work.
Anthropic frames the change as defender-aligned: faster public knowledge when vendors stall. Maintainers may experience it as pressure—the post acknowledges that and promises actionable guidance, patches or mitigations where possible, and paced report volume to overloaded projects.
Maintainer-facing programs Anthropic cites
Anthropic points maintainers to three programs in the same announcement:
- OSS Scanner — opt-in periodic model scans; Anthropic describes unreviewed model output in its launch post.
- Claude for Open Source Program — six months of Claude Ultra at no cost for eligible maintainers, named in Anthropic's CVD update.
- Cyber Verification Program — tailored safeguards for legitimate defensive work, also named in the same CVD update.
None of these replace your patch pipeline; they explain where new findings may originate before they hit public disclosure on a shorter clock.
Backlog disclosures starting January 2027
Anthropic says it will contact maintainers of high and critical issues reported more than 90 days ago that still lack public disclosure, grant an additional 30 days, then publish on its dashboard aligned with the January 1, 2027 policy. Maintainers should watch for that mail if they ever received a private Anthropic report and assumed the clock was indefinite.
Agent users should not wait for the dashboard: if you depend on a package named in community chatter about impending Anthropic publication, treat that as a patch-soon signal even before CVE text is public.
Exceptions Anthropic explicitly keeps
The announcement names cases where 30 or 45 days is unrealistic: critical infrastructure deployments, specification or RFC changes, and coordinated vendor extensions Anthropic negotiates case by case. The seven-day rule for active exploitation is unchanged—if you run agents against internet-facing services, KEV and vendor exploit flags still outrank calendar disclosure.
Human review remains the gate for unreviewed OSS Scanner output entering CVD timelines, which is Anthropic's answer to false-positive floods from autonomous scanners—similar to the non-functional PHP2Shell PoCs Google GTIG documented overwhelming defenders.
What this means if you run AI coding agents
Coding agents accelerate how fast vulnerable dependencies land in production, not Anthropic's legal relationship with upstream maintainers. Practical checks:
-
Shrink merge windows for security bumps. If public detail may arrive 30 days after report instead of 90, treat agent-proposed
package.jsonor lockfile updates as release blockers when scanners flag known CVEs. -
Subscribe to vendor and coordinator feeds. Track Anthropic advisories, GitHub Security Advisories, and distro notices for stacks agents touch (Node, Python, Rust MCP crates).
-
Do not rely on quiet periods for secrets. Shorter CVD does not excuse leaving credentials in repos agents read; see gitee live keys research and GitGuardian's parallel enterprise disclosure push in our November 16 coverage.
-
Separate scanner noise from reviewed findings. OSS Scanner and similar tools may flood maintainers; Anthropic excludes unreviewed AI output from the new timeline rule—your triage process should too.
-
Align with other disclosure frameworks. Compare Anthropic's move to OpenAI's Patch the Planet stages—multiple labs are compressing OODA loops; your incident playbooks should assume hours-to-days weaponization for internet-facing deps.
-
Log agent dependency edits. When a disclosure lands, you will want to know whether an agent merged the vulnerable version; local observability helps.
-
Rehearse coordinator handoffs. If you maintain a widely used library, practice what happens when Anthropic escalates to an external coordinator at day 15—that is sooner than many legal teams expect.
How this interacts with other labs' disclosure rhetoric
Anthropic is not alone in compressing timelines. OpenAI's Patch the Planet framework discusses tiered early notice and asks when authorized parties might remotely patch systems—far more aggressive than publication alone. Google's GTIG vulnerability trends post documents n-day weaponization rising with LLM assistance.
For a maintainer, the practical merge is: assume parallel discovery, ship fixes faster than your old CVD comfort zone, and communicate with coordinators when RFC-level work is required. For an agent user, the merge is: your lockfile is on the same clock as the maintainer's embargo.
January 2027 checklist for AppSec leads
| Week | Action |
|---|---|
| Now | Inventory dependencies Anthropic or other labs have privately reported on |
| Q4 2026 | Add same-day severity triage for agent-opened Dependabot or SCA tickets |
| January 2027 | Re-read Anthropic's public dashboard when backlog publications begin |
| Ongoing | Document whether agents may bump transitive deps without human ack |
None of this requires changing Claude Code settings on day one; it requires organizational SLAs that match the new publication speeds.
What Anthropic's coordinated vulnerability disclosure policy means for searchers
People querying anthropic vulnerability disclosure or anthropic responsible disclosure policy are usually deciding whether a bug they found will be published on a fixed schedule. Anthropic's November 16 post answers: yes, and the schedule is shorter for severity starting in 2027. The post also clarifies that OSS Scanner output without human review does not automatically enter that schedule—maintainers enrolling in OSS Scanner choose faster unreviewed reports separately from CVD timelines.
If you file bugs through Anthropic as a researcher, re-read the coordinator escalation at day 15 rule: silence is no longer a month-long pause.
Coding-agent vendors and hook authors should also expect more public write-ups on Claude-adjacent tooling bugs on the shorter clock, even when the flaw lives in an npm dependency rather than in Claude itself. Plan embargo comms accordingly.
What Beam can and cannot do here
Beam does not participate in coordinated disclosure and does not patch upstream packages. It helps on the workstation side:
- PreToolUse deny (Claude Code and Codex, where wired) can block destructive upgrade commands while you verify an advisory.
- Heuristic skill/MCP scans catch risky install patterns before agents pull unaudited proof-of-concept code when a shortened disclosure drops.
Beam remains heuristic, hook-based, and fail-open without cached policy. It is not a substitute for dependency bots, SBOM review, or maintainer patch SLAs.
Frequently asked questions
When do Anthropic's new CVD timelines take effect?
They apply to vulnerabilities Anthropic reports to vendors and maintainers on or after January 1, 2027. Reports already in progress as of that date keep their existing deadlines.
What are the new disclosure windows?
High and critical severity move from 90 days to 30 days. All other severities move from 90 days to 45 days, per Anthropic's November 16, 2026 announcement.
What is Anthropic's coordinated vulnerability disclosure policy?
Anthropic's CVD policy sets deadlines for public disclosure after it privately reports bugs to maintainers. The November 16 update shortens those deadlines and adjusts escalation and post-patch detail timing because AI-assisted discovery increases rediscovery risk.
Does the 30-day window apply to OSS Scanner findings automatically?
No. Anthropic states the update does not apply to AI-discovered findings that have not received human review, such as those from OSS Scanner.
What happens if a maintainer does not respond?
Anthropic will aim to escalate to an external vulnerability coordinator after 15 days instead of 30 when the maintainer has not responded to the initial report.
When does Anthropic publish full technical details after a patch?
It will wait 30 days after a patch is available before sharing full technical details, down from 45 days previously.
Are vulnerabilities under active exploitation treated differently?
Yes. Anthropic's seven-day disclosure window for any vulnerability under active exploitation remains unchanged.
What should coding agent users do before January 2027?
Shorten dependency upgrade SLAs, watch Anthropic's published advisories for libraries your agents bundle, and treat model-assisted scanner noise as a signal to patch sooner—not as a reason to defer.
Related reading
- CVE volume record in 2026
- OpenAI Patch the Planet framework
- Anthropic AI-built exploit chains
- GTIG exploit_agent and n-day pressure
- Google Project Zero Full X disclosure schedule
- PHP2Shell and agent-assisted patch urgency
Anthropic's announcement was reviewed on November 16, 2026 UTC. Link OSS Scanner and program URLs to Anthropic's site for current eligibility rules.

