Google Project Zero ended its "90 + 30" vulnerability disclosure rule on November 13, 2026, and replaced it with a staged "Full X" policy: public technical details now appear X days after report to a vendor—even if no fix has shipped—with X at 90 today, 60 from February 2, 2027, and 30 from January 3, 2028. The 2026 policy post by Natalie Silvanovich cites AI-assisted discovery, faster rediscovery, and near-immediate patch-to-exploit work as reasons the old soak period no longer matches attacker behavior.
If you run Claude Code, Codex, Cursor, or similar agents on a laptop that shares Chrome, .NET, Qualcomm firmware, and kernel packages with Project Zero's targets, this is not a distant standards debate—it is a clock on when full write-ups and repro steps may hit the public internet while your agents still automate installs and refactors against those stacks.
| Question | Short answer |
|---|---|
| Am I affected? | Indirectly, yes if your dev machine or CI images run software Project Zero reports on (browsers, runtimes, kernels, drivers) |
| Is there a fix? | No product patch—change how fast you apply vendor updates and how you triage disclosure feeds |
| Does data leave my machine? | Not from this policy alone; faster public detail can speed third-party exploit tooling against unpatched hosts your agents touch |
| When did it start? | November 13, 2026 (Full 90); February 2, 2027 (Full 60); January 3, 2028 (Full 30) |
| What disappeared? | The extra 30-day soak after vendors patched under the old 90+30 model |
explainx.ai summarized Anthropic's same-day CVD compression for readers who want a news timeline across both announcements.
What Project Zero changed (mechanism)
Since 2021, Project Zero gave vendors up to 90 days to remediate, then another 30 days of "soak time" so patches could roll out before publication. Silvanovich writes that goal made sense when patch gaps were the main user-risk window.
The November 2026 announcement argues the landscape shifted on three axes relevant to agent operators:
- Discovery speed — Large language models help attackers find and exploit flaws faster, and increase the chance multiple parties discover the same bug during embargo.
- Patch differential speed — Attackers can reverse patches quickly enough that "zero patch-to-exploitation time" is a reasonable model; soak time can create a false sense of secrecy after a fix ships but before users install it.
- Report volume — Project Zero now finds several vulnerabilities per week in targets that once yielded a few per year, making per-bug negotiation over grace periods and variant handling "quickly becoming unmanageable."
The replacement policy, "Full X," is deliberately simple for automation:
| Phase | Effective date | Rule |
|---|---|---|
| Full 90 | November 13, 2026 | Publish 90 days after report to the vendor |
| Full 60 | February 2, 2027 | Publish 60 days after report |
| Full 30 | January 3, 2028 | Publish 30 days after report |
The permanent policy page (projectzero.google/vulnerability-disclosure-policy.html) illustrates behavior with examples: if a vendor patches on day 47, details still go public on day 90 under Full 90; if they never patch, details still go public on day 90.
Exceptions that survived the simplification
Project Zero removed most special cases—including grace period and variant policy—but kept:
- In-the-wild exploitation — Handled under Google's application security disclosure policy with a seven-day public timeline when active exploitation against real users is confirmed.
- Already-public flaws — Project Zero may publish immediately when someone else has disclosed details.
- Vendor-requested early disclosure — Vendors can ask Project Zero to publish ahead of the deadline.
- Rare harm exceptions — Silvanovich notes Spectre/Meltdown-style cases may still get longer withholding; only five exceptions were granted in twelve years.
Severity is not a separate disclosure knob. Project Zero argues severity affects user impact if exploited, not the odds an attacker finds the bug during embargo—so the calendar does not lengthen for "lower" issues once they enter the tracker.
Why Google staged the drop instead of jumping to 30 days immediately
Google solicited feedback from vendors that receive the most Project Zero reports. Vendors worried an instant 30-day rule could publish unfixed bugs before remediation pipelines matured, encourage priority inversion (rushing Project Zero tickets over worse bugs reported elsewhere), or create learned helplessness if maintainers believe they cannot meet the deadline anyway.
Google's response is the gradual Full 90 → 60 → 30 schedule—"challenging but realistic"—with the soak extension removed immediately because it no longer matches patch-gap risk.
How Full X interacts with AI-era exploit economics
Project Zero's reasoning rhymes with separate research Beam has covered on patch gaps and autonomous patch-to-exploit pipelines:
- Epoch AI's Firefox patch-gap study showed open-weight models building file-read exploits from public but not yet shipped fixes—about $1,000 per version in their accounting.
- Google GTIG's exploit_agent demo went further: an agentic workflow from patch diff toward working exploit material, with GTIG framing it as a defensive red-team tool—but the mechanism is what attackers also buy down.
Put together, Full X tells maintainers: users may remain exposed after your patch ships, and embargo may not hide the bug from parallel discoverers. Project Zero will not wait for you to win a debate over whether a non-patch mitigation "counts" as fixed before the calendar fires—though they will still communicate fix status.
For volume context, see Beam's CVE count explainer: more IDs do not equal more exploitation, but shorter disclosure clocks mean fewer weeks to hide behind "we are still rolling out."
Timeline comparison: Project Zero vs Anthropic CVD (same week)
Google announced Full X on November 13, 2026; Anthropic published its shorter CVD windows on November 16, 2026. Both cite AI-accelerated discovery risk, but they govern different reporters:
| Program | Who reports | Near-term change | Long-term floor |
|---|---|---|---|
| Project Zero Full X | Google's research team to any vendor | Full 90 from Nov 13, 2026 (no +30 soak) | Full 30 from Jan 3, 2028 |
| Anthropic CVD | Anthropic to OSS maintainers for its findings | 30 days high/critical, 45 days other severities for reports on or after Jan 1, 2027 | Maintains 7-day rule for active exploitation |
Coding teams should track both feeds if agents pull from ecosystems both programs scan—Anthropic's OSS Scanner noise is explicitly out of scope for the new CVD clock until human review, while Project Zero's tracker entries are already curated for demonstrated impact.
What this means if you run AI coding agents
Agents do not patch your OS, but they amplify exposure when they:
- Install or upgrade npm/pip/cargo packages on schedules slower than disclosure calendars.
- Drive browser-based MCP tools against Chrome-family engines Project Zero regularly audits.
- Script CI images that lag laptop patches by weeks.
Concrete habits that match Google's stated goals (faster remediation, smaller patch gap):
- Match SLAs to Full X, not to "Patch Tuesday plus a month." If a dependency is in a stack Project Zero has historically reported on—think Chromium, Windows kernel surfaces, .NET, Qualcomm—assume technical detail on day 90 (soon 60, then 30) even when your distro lags.
- Prioritize KEV and EPSS before CVSS alone. Project Zero's volume rise is one input; CISA KEV and EPSS still separate exploited-in-the-wild risk from advisory noise—agents should not auto-bump every transitive dependency because a CVE count spiked.
- Watch agent-proposed upgrades skeptically. Models may suggest
apt upgradeor bulknpm updateafter reading headlines. Confirm vendor advisories and your change window before letting an agent rewrite lockfiles on production branches—see PHP2Shell timing lessons on how fast weaponization moved in 2026. - Separate research hosts from daily driver. If you experiment with bug-hunting skills or GTIG-style tooling, run them on disposable VMs whose patch level you control; do not mix with tokens your agent stores for production repos.
Checking versions (examples)
Project Zero does not publish a single "agent bill of materials," but these checks mirror stacks its recent advisories touched:
google-chrome --version
dotnet --list-runtimes
uname -r
npm ls --depth=0
Compare against vendor security pages the same week Project Zero's reporting transparency lists a new entry for your product.
How to watch the clock without drowning in CVE noise
Project Zero continues to list vendor, product, and report date on its reporting transparency page soon after notification, as described in its disclosure policy page. That is the earliest honest signal that a fixed-calendar countdown started—even though technical detail stays private until day X.
For coding-agent workflows, treat that signal differently from a generic NVD drop:
- Lockfile repos: If transparency names a library your agent vendors into Docker images, open a ticket immediately; do not wait for CVSS enrichment when NVD backlog is thin (context).
- Browser MCP tools: Chromium-family advisories often move in clusters; when Project Zero's target list overlaps your default automation browser, schedule an upgrade window before day 90 under current Full X.
- Distro lag: Full X publication can arrive before Ubuntu LTS or enterprise macOS channels ship backports; plan a user-space mitigation (disable feature, container pin) instead of assuming embargo protected you.
None of this replaces vendor patches—it aligns agent automation with a disclosure program that explicitly stopped rewarding slow rollout.
What Beam can and cannot do here
Beam does not shorten disclosure deadlines or patch your dependencies. It records and optionally denies risky agent actions on-device when workspace policy is in enforce mode for supported hooks (Claude Code and Codex PreToolUse).
What helps in a faster disclosure era:
- Audit trails of package installs,
curl | bashpatterns, and credential-bearing tool calls—useful when you are proving whether an agent upgraded a library before or after a public PoC dropped. - Heuristic skill/MCP scans before you enable new automation that might pull exploit code from fresh write-ups.
Limits to state plainly: Beam is not a kernel sandbox or vulnerability scanner; missing cached policy is fail-open; Cursor and other harnesses may only get advisory notes unless you wire equivalent controls.
Related reading
- Anthropic CVD timelines from January 2027
- CVE volume record and triage habits
- Google GTIG exploit_agent and patch differential risk
- Firefox patch-gap exploits with open-weight models
- PHP2Shell and agent-driven exploit speed
Primary sources: Project Zero — Policy and Disclosure: 2026 Edition, Project Zero — Vulnerability Disclosure Policy.
Frequently asked questions
What is Google Project Zero's new disclosure policy called?
Project Zero calls it "Full X." Technical details publish X days after report, independent of patch status. X = 90 starting November 13, 2026.
When does Project Zero move to a 30-day disclosure window?
January 3, 2028, under the published schedule. February 2, 2027 is the intermediate Full 60 step.
Did Project Zero remove the 30-day patch rollout extension?
Yes. The 90 + 30 soak model ended November 13, 2026. There is no automatic extra month after vendors ship fixes.
Will Project Zero publish unfixed vulnerabilities?
Yes. After the Full X deadline, Project Zero discloses even without a vendor patch, unless a rare exception applies.
What exceptions remain under Full X?
Seven-day handling for confirmed in-the-wild exploitation, immediate publication when others already disclosed, vendor-requested early release, and rare user-harm exceptions documented on the policy pages.
How does Full X relate to Anthropic's new CVD windows?
They are parallel programs from different reporters, both citing AI-accelerated risk. Track Anthropic's maintainer deadlines separately in /blog/anthropic-cvd-disclosure-30-day-window-january-2027.
Does Project Zero still negotiate whether a bug is fixed before disclosure?
They still track and communicate fix status, but fix verification no longer delays publication under Full X.
What should coding agent users do after this policy change?
Shorten upgrade cycles, prioritize KEV/EPSS, verify agent-proposed bulk updates, and read vendor advisories for browsers, runtimes, and drivers your agents automate.
Policy dates and quoted timelines are taken from Google's November 13, 2026 Project Zero posts and were accurate as of publication on November 16, 2026.

