Policy#Security scan#AI safety

Project Zero drops the 30-day grace period: bugs go public 90 days after report

Project Zero now publishes bugs 90 days after vendors get the report, dropping the 30-day grace period; the window falls to 60 days in 2027 and 30 in 2028.

A mechanical stopwatch face on a white background, with a red 60 mark and black hands

Project Zero published “Policy and Disclosure: 2026 Edition” on November 13, replacing the “90 + 30” policy it had run since 2021 with “Full X”: a reported vulnerability becomes public X days after the vendor receives the report, whether or not a patch exists, and X is 90 for now. The schedule tightens to 60 days on February 2, 2027 and to 30 days on January 3, 2028. This is the disclosure clock for bugs Project Zero reports to vendors — Google’s own product teams sit in the same queue — not a new rule that Google imposes on the industry. The Hacker News thread (story 50496051) was posted on November 14 and stood at 220 points and 78 comments at the time of writing, and most of the argument was about the 30-day endpoint in 2028.

Key points

  • The 90 days count from the day the report reaches the vendor, not from the day a patch ships. The official policy page spells out three cases: patched on day 47, patched on day 90, never patched — all three go public on day 90.
  • The old 90-day remediation window and the 30-day “soak” for patch rollout are both gone, along with the grace period and the variant rule. Disclosure no longer depends on whether Project Zero and the vendor agree that a bug is fixed.
  • Three exceptions remain: vulnerabilities exploited in the wild follow Google’s application security disclosure policy, which publishes after 7 days; details already published by someone else can go out immediately; and a vendor that asks for early disclosure gets it. Exceptions overall are rare — five granted in 12 years — and hardware cases such as Spectre and Meltdown are among them.

Background

The policy that just ended gave vendors 90 days to remediate plus 30 days of soak time to get the patch to users, so disclosure could land as late as day 120. Its premise was that the extra month helped users get fixes sooner. The blog post rejects that premise directly: LLMs have cut the time it takes an attacker to reverse-engineer a patch and write an exploit to the point where “zero patch-to-exploitation time is a reasonable, though imperfect, model,” and sitting on details in that world creates “a false sense of secrecy.”

The groundwork for the change was already visible in October. On October 27, Project Zero published a comparison of its JavaScript engine fuzzer Fuzzilli against BigSleep, its LLM-based bug finder: over six months BigSleep reported 45 vulnerabilities in V8 against 37 for Fuzzilli, with measured overlap of 25% and an estimated theoretical overlap of about 60% — meaning a substantial share of bugs were findable by only one of the two approaches. On October 6 the team published a note on the fast remediation channels large vendors keep for urgent fixes, which reads in hindsight like advance warning that vendors would need one. Add the blog post’s own volume line — “several vulnerabilities a week in targets where we once found a few vulnerabilities a year” — and the argument is that discovery scaled while delivery did not.

For comparison, ZDI (TrendAI’s Zero Day Initiative) still runs a 120-day standard window as of November 15, 2026, with extensions granted only in rare circumstances. Project Zero’s 90 days is three quarters of that window; the 30-day target is a quarter.

The facts

The policy defines X as the window in which a vendor is expected to make a patch available to users, and disclosure fires at the deadline regardless. That distinction is the whole story: vendors used to be able to trade “we’re close to a fix” for more time, and now the only thing that buys time is the vendor asking for early publication. The project also stopped varying the clock by severity. Its reasoning: severity does not change whether attackers can find a bug, only what happens when they do, so vendors using severity to schedule fixes are deciding which risks their users absorb.

PolicyRemediation windowGoes publicGrace periodExceptions
Old (from 2021)90 daysby day 12030 daysnegotiated case by case
Full 90 (2026-11-13)90 daysday 90noneITW after 7 days, already public, vendor request
Full 60 (2027-02-02)60 daysday 60nonesame
Full 30 (2028-01-03)30 daysday 30nonesame
ZDI (for comparison)120 daysday 120rarefaulty patches on a 30/60/90-day tier

The policy in practice is readable from Project Zero’s Reporting Transparency page, a trial that began on July 29, 2025 and lists the vendor, product, report date and deadline for each report. At the November 10, 2026 update the page carried 115 rows, with report dates running from January 7 to November 6; by recipient that is Apple 44, Adobe 16, Google 15, Samsung Mobile 12. Two rows check out by hand: PZ-582900381, a Pixel Wave6 media codec bug reported to Google on 2026-NOV-06, expires 2027-FEB-04; PZ-580895393, an Adobe DNG SDK bug reported 2026-NOV-02, expires 2027-JAN-31. Both are exactly 90 days after the report date, and Google’s own hardware is in the same line as everyone else.

What others say

On Hacker News the objection was less about the direction than about the endpoint. Wowfunhappy accepted the case for deleting the 30-day rollout buffer but called a 30-day total “needlessly harsh” and unsupported by the post: “What if the bug is actually inherently hard to fix?” sippingabonedry added the operational version — “It’s not the fixing that’s usually the problem, it’s more the testing,” and skipping it is how you get another CrowdStrike. kapilvt asked what happens when the bug is in a widely used library baked into embedded devices; cookiengineer noted the irony that AOSP still holds patches back for six months. Others read the move as leverage: bmitc asked whether it is “a version of extortion,” and sethaurus answered that Project Zero does not create vulnerabilities, it finds them, that vendors want unlimited time while users want to know, and that the tension is why a deadline exists at all. ares623’s joke — a “Full 0” policy in 2029 — became the thread’s running gag.

Vendors get a paragraph in the post but no names. Project Zero says it consulted “the vendors we report the most vulnerabilities to,” and relays three concerns: that switching immediately would expose unfixed bugs while remediation processes are still maturing; that an aggressive deadline would push vendors to prioritize newly reported bugs over more serious older ones; and that enforcing a schedule some vendors cannot meet would lead them to conclude they will never finish and stop trying. Those three concerns are the stated reason the change is staged rather than applied at once.

One gap in the record: as of November 15, 2026, searches turned up no independent coverage from The Record, BleepingComputer, SecurityWeek or CyberScoop. Reaction is concentrated on Hacker News, so there is no third-party tally of whether 30 days is workable — only vendor paraphrases and practitioners’ judgement.

Our take

For vendors and open-source maintainers, what changed is not the fix deadline but the publication date. The clock starts when the report lands, so remediation SLAs have to be built backwards from a disclosure date rather than forwards from a release train. The hard part is distribution and rollback: between a merged patch and an updated device sit testing, staged rollout, store review and enterprise change windows, and 30 days does not fit most of that. That is the strongest objection in the thread, and the policy does not answer it beyond saying vendors should prepare non-patch mitigations.

Do not count on the exceptions. The policy page is explicit that they are “rare and publicly recorded” — five in 12 years — and the examples the blog gives are Spectre and Meltdown, infrastructure-scale problems. A flagship product with one high-severity bug does not clear that bar. Project Zero also keeps a self-imposed filter: it only reports vulnerabilities with demonstrated security impact that it can reproduce, which moves the screening to before the report. The takeaway is not that every bug now surfaces in 30 days, but that anything in this tracker is a bug Project Zero considers serious.

Two inconsistencies are worth recording. The blog says the three listed cases are “the only exceptions,” while the same post also has a “What about hardware?” section acknowledging that exceptions will continue to be granted, and the policy page adds discretion for “wild-scale user harm.” And the blog says the project “will move” to 60 and 30 days, while the policy page says “we plan to.” One is a commitment, one is a plan, and whether February 2027 slips is unanswered.

The practical reference is the transparency page, updated weekly: it shows which vendors ship inside 90 days and which run to the deadline. Our own recent entries bracket the range — OpenSSH’s musl authentication bypass shipped an emergency release the day after the report, while OpenAI’s autonomous pipeline shows how fast discovery has become. Put those two ends together and the 30-day number stops looking arbitrary.

What to watch

Two dates matter: February 2, 2027 for Full 60 and January 3, 2028 for Full 30. Until then, 90 days is a hard line. The way to test enforcement is not to reread the policy but to compare deadlines against actual publication dates on the transparency page each week.