Anthropic and Google drop the 90-day bug deadline: who pays for 30-day disclosure in the AI era
Anthropic and Google Project Zero are cutting bug-disclosure deadlines from 90 to 30 days, citing AI. With thousands of AI-found reports, maintainers and patch teams carry the load.
Within four days in November 2026, Google Project Zero and Anthropic's Frontier Red Team both scrapped the 90-day disclosure deadline that has been the industry default for more than a decade. Both say AI has made bugs faster to find, faster to rediscover and faster to exploit from a patch. Anthropic alone has reported 5,087 AI-found vulnerabilities to open-source projects, of which it knows of 629 that have been patched — so in practice, details of many serious bugs will be published while the software is still unfixed or not yet updated.
The short version
When a security researcher finds a bug, the usual practice is "coordinated disclosure": tell the vendor or open-source maintainer privately, give them time to fix it, then publish the details. For years the default clock was 90 days. Google Project Zero added an extra 30 days of "soak time" after a patch so users could install it before the details came out.
On November 13, 2026, Project Zero replaced that with a flat rule it calls "Full X": details go public X days after the report, patched or not. X is 90 now, 60 from February 2, 2027 and 30 from January 3, 2028. On November 16, Anthropic announced that, for bugs reported from January 1, 2027, it will publish high- and critical-severity findings after 30 days and all others after 45, and will wait 30 days (down from 45) after a patch before releasing full technical details.
Both organisations give the same reasons. AI tools now find bugs at a rate humans cannot match, different models increasingly find the same bug independently, and attackers can use AI to work out the flaw behind a patch almost immediately. Anthropic is a large source of such reports: its dashboard counts 5,087 vulnerabilities reported across 680 open-source projects. The open question is who absorbs the shorter clock: volunteer maintainers, vendors with monthly release trains, and the IT teams who install updates.
Timeline
- 2021 — Google Project Zero adopts its "90 + 30" policy (90 days to fix, up to 30 more days after a fix before details are published), as described in its November 2026 post.
- March 6, 2026 — Last update of Anthropic's coordinated vulnerability disclosure (CVD) policy: 90-day deadline, a 45-day wait after a patch, escalation to an outside coordinator after 30 days of silence.
- June 9, 2026 — VulnCheck's Patrick Garrity analyses Anthropic's disclosure ledger: 1,596 findings reported, 27 shown as fixed, 10 undisclosed findings already past 90 days.
- October 20, 2026 — curl announces that from October 26 low- and medium-severity vulnerability reports will be handled publicly on GitHub, citing an "avalanche" of reports (The Register).
- November 2, 2026 — Last update of Anthropic's discovery dashboard: 5,087 vulnerabilities reported to maintainers, 629 known to be patched.
- November 13, 2026 — Project Zero publishes Policy and Disclosure: 2026 Edition by Natalie Silvanovich, introducing "Full X". The post-patch grace period and variant policy are removed.
- November 16, 2026 — Anthropic's Frontier Red Team publishes Updating our coordinated vulnerability disclosure policy timelines. The same day, Google's threat intelligence group publishes its Q4 AI Threat Tracker describing an attacker framework that turns patches into exploits automatically (our coverage).
- January 1, 2027 — Anthropic's new deadlines apply to newly reported bugs; overdue high/critical reports start being published on its dashboard after a final 30-day notice.
- February 2, 2027 — Project Zero moves to Full 60.
- January 3, 2028 — Project Zero moves to Full 30.
What we know
Project Zero's new rule. According to Project Zero, vulnerabilities are disclosed X days after being reported to the vendor, and fix status "will no longer be a decision point". The team will still track whether bugs get fixed and say when they have not. Exceptions remain for bugs exploited in the wild (handled under Google's broader disclosure policy), bugs already made public by others, vendors who ask for earlier disclosure, and rare hardware cases such as Spectre and Meltdown. Project Zero says the reduction is gradual because vendors warned that an immediate switch could expose unfixed bugs, and could push them to prioritise by deadline instead of severity or to give up on meeting deadlines. It says it now finds "several vulnerabilities a week" in targets where it once found a few a year, and that post-patch soak time "can put users at risk by creating a false sense of secrecy". We covered the policy itself in a shorter piece on November 16.
Anthropic's new rule. Anthropic's November 16 post sets:
- 30 days for high- and critical-severity issues, 45 days for everything else (previously 90 for all);
- escalation to an external vulnerability coordinator after 15 days without a maintainer response (previously 30);
- full technical details 30 days after a patch is available (previously 45);
- the seven-day window for actively exploited bugs unchanged;
- negotiated timelines for hard-to-deploy software such as critical infrastructure, and for flaws that need a specification or RFC change.
The new deadlines apply to reports made on or after January 1, 2027; reports already in progress keep their deadlines. High- and critical-severity reports filed more than 90 days ago and still undisclosed get a final 30 days, after which Anthropic says it will "publish these reports on our dashboard" under the updated policy from January 1, 2027. The update "will not apply to AI-discovered findings that have not received human review", such as those from its free OSS Scanner service for open-source maintainers.
Anthropic's stated reason: the old policies assumed "fixes moved at human speed" and that "independent rediscovery of the same bug was unlikely". "Neither assumption holds today," it writes, adding that it is "seeing a higher rate of vulnerability rediscovery" and that "a bug that one model can find is increasingly likely to be found by another." The post gives no figures for that rediscovery rate.
The scale behind it. Anthropic's dashboard, last updated November 2, 2026, lists 37,966 candidate findings, 6,376 reviewed by external security firms, and 5,087 vulnerabilities reported to maintainers across 680 open-source projects. Of those, 3,635 were acknowledged by maintainers and, "to our knowledge, 629 of these have been patched". It counts 733 identifiers issued (230 CVE records and 503 GitHub Security Advisories). In June, VulnCheck counted 1,596 reported findings and 27 fixed on Anthropic's ledger at the time, and noted that 10 undisclosed findings were already past the 90-day window, with 168 more due within 30 days.
The maintainer side. curl's security policy now asks reporters of LOW or MEDIUM issues to "submit it publicly" on GitHub; only HIGH and CRITICAL stay private on HackerOne. Maintainer Daniel Stenberg told The Register: "We must act to survive." curl closed its bug bounty earlier in 2026 because of AI-generated submissions. Next.ink reports that researcher Marcus Hutchins criticised the Google change on X, arguing that overworked maintainers already fix bugs as fast as they can, and that VideoLAN said deploying fixes is as much of a problem as producing them. We have not seen those posts directly.
Technical analysis
Two problems that 90 + 30 used to solve. A disclosure deadline balances two risks. Before a patch exists, publishing helps attackers more than defenders. After a patch exists but before people install it, publishing the details again helps attackers, unless they could work them out anyway. The 90-day clock addressed the first risk by pushing vendors to fix; the soak time addressed the second by keeping details private while patches spread.
AI attacks both assumptions. On the first, if another team or model is likely to find the same bug soon, a long private window does not keep the bug secret; it only keeps defenders unaware. On the second, soak time works only if a patch hides its bug. Google's own threat intelligence team now reports an attacker framework that watches for new releases, compares old and new code, writes an exploit and runs it without a human in the loop. If that is typical, the patch itself is the disclosure, and holding back the write-up protects very little.
The two policies differ in useful ways.
- Severity. Anthropic gives high and critical bugs the shorter window (30 days) and others 45. Project Zero applies one clock to everything and says severity is not an effective disclosure metric because it "doesn't change how likely attackers are to find a bug". Both reverse the old intuition that serious bugs deserve more time to fix carefully. The logic is that serious bugs are the ones attackers are most likely to be hunting.
- Post-patch wait. Project Zero drops it entirely. Anthropic keeps a shorter one: 30 days after a patch before "full technical details". So Anthropic still believes some embargo is worth having; Project Zero has concluded it is not, except for the transition years.
- Speed of change. Project Zero reaches 30 days in January 2028 after a 60-day stage. Anthropic reaches 30 days for its most serious findings on January 1, 2027, six weeks from now, with no intermediate step.
The arithmetic of AI-scale reporting. Our assessment: the real pressure comes from volume, not the deadline itself. A human research team reporting a few dozen bugs a year to large vendors can live with almost any clock. Anthropic's dashboard shows 5,087 reports across 680 projects, with 629 known patches. That is about 12% by Anthropic's own count, which may understate fixes it has not tracked. Even allowing for that, a 30-day clock on findings of this volume means many will reach their deadline unfixed. The overdue clause makes this concrete: high- and critical-severity reports older than 90 days will start appearing on the dashboard in early 2027.
There is a real tension between the two arguments both organisations make. If rediscovery is common, publishing quickly mainly removes a false sense of secrecy. But the same AI tools that find bugs also read published details, so a public write-up of an unpatched bug in a small, widely embedded library could give attackers something new. The policies bet that the first effect outweighs the second. Neither post offers data to show that, and Project Zero describes zero patch-to-exploit time as a model, not a measurement.
What is new about the AI angle. Fixed deadlines are not new, and Project Zero has argued for shorter ones since at least 2021. What is new is that the organisations shortening the deadlines are also among the largest producers of AI-found vulnerability reports. Anthropic sells the models that generate many of these findings and offers free scanning and subscriptions to maintainers as mitigation. That does not make the policy wrong, but it means the actor setting the pace is also one of the main sources of the load.
What remains unclear
- What is published at a deadline when no patch exists. Anthropic's post specifies when full technical details follow a patch, but not what it publishes when the 30- or 45-day deadline passes without one. Its policy page, still dated March 6, 2026 on November 19, says details go public at the deadline "unless there's a compelling security reason" otherwise.
- How severity is decided. Anthropic's tiers depend on severity, but its dashboard shows its ratings agree exactly with external firms in 88.0% of 1,868 cases. VulnCheck noted in June that Claude's scores ran higher than third-party assessments. Who has the final say when a maintainer disagrees is not stated.
- Rediscovery data. Both organisations say independent rediscovery is rising; neither publishes a rate.
- Fix status. The 629 patched figure is labelled "to our knowledge", so the real fix rate may be higher. VulnCheck saw patches that appeared to predate Anthropic's reports, which suggests other researchers had already found some of the same bugs.
- How many reports are already overdue. Anthropic has not said how many high/critical reports fall under the 90-day overdue clause, or which projects will be affected.
- Whether others follow. We found no comparable change from bug-bounty platforms, national CERTs or other AI labs as of November 19, 2026.
- Monthly patch cycles. A 30-day clock sits awkwardly with vendors that ship security fixes monthly; neither post addresses that directly beyond Project Zero's advice to plan for mitigations that do not depend on a patch.
Lessons and what to do
For security and IT operations teams
- Assume details of serious bugs in common software will be public within weeks of the vendor learning about them, and sometimes before a patch exists. Build an emergency patch path measured in days for internet-facing and widely deployed software.
- Track open-source dependencies, not only vendor products. Much of Anthropic's reporting goes to open-source projects that end up inside appliances and commercial software. A software bill of materials (SBOM) and a dependency scanner tell you which advisories matter to you.
- Prepare non-patch mitigations in advance: feature flags, network restrictions, WAF rules and monitoring for systems you cannot patch quickly.
- Add Anthropic's dashboard and Project Zero's tracker to your watch list from January 2027, alongside vendor advisories.
For open-source maintainers and product security teams
- Publish a working security contact and a SECURITY.md. Anthropic will escalate to an outside coordinator after 15 days of silence from January.
- If you receive AI-generated reports, use the pacing Anthropic says it offers: agree a sustainable rate before large batches arrive. Ask for the human-review status of each report and for a candidate patch where available.
- Consider curl's approach: handle lower-severity issues in public so more contributors can help, and keep private handling for high and critical.
- Vendors with monthly release cycles should decide now how they will ship out-of-band fixes for bugs on a 30-day clock.
For AI builders running vulnerability discovery
- Volume is now the binding constraint. Validate before reporting, bundle fixes with reports, and measure fix rates per project, not just reports sent.
- Publish rediscovery and fix-rate data. The case for short deadlines rests on rediscovery; data would make it stronger and help maintainers plan.
For leadership
- Revisit patch SLAs written for a 90-day world. "30 days for high, 90 for medium" assumes a quiet period that no longer exists.
- Budget for open-source maintenance you depend on. If AI-found reports keep growing faster than maintainers can fix them, unpaid projects in your supply chain become your problem.
Sources
- Anthropic Frontier Red Team: Updating our coordinated vulnerability disclosure policy timelines (November 16, 2026)
- Anthropic: Coordinated vulnerability disclosure for Claude-discovered vulnerabilities (policy page, last updated March 6, 2026)
- Anthropic: vulnerability discovery dashboard (last updated November 2, 2026)
- Google Project Zero: Policy and Disclosure: 2026 Edition (November 13, 2026)
- curl: Vulnerability Disclosure Policy
- VulnCheck: Observations on Anthropic's Vulnerability Disclosure Ledger (June 9, 2026)
- Coverage: Help Net Security (November 17, 2026), Next.ink (November 17, 2026), SecurityWeek, The Register (October 20, 2026)
- Related on AI Cyber Brief: Project Zero cuts its bug deadline to 30 days by 2028; Google: AI agents now turn security patches into working exploits