Project Zero ends 30-day soak time and keeps 90-day disclosure window

Google’s Project Zero will end its 30-day soak time on Nov. 13, 2026. Starting that day, the group will disclose flaws 90 days after a vendor gets a report under its “Full 90” policy.

Since 2021, Project Zero operated under a “90 + 30” policy that allowed vendors as many as 90 days to fix a flaw and another 30 days of soak time. Project Zero said the older approach aimed to speed up patching and give vendors time to ship fixes to users. Google Project Zero made the 90-day disclosure window popular in 2014 to create a predictable timeline for vendors and researchers.

Project Zero says AI-assisted discovery and exploitation have shifted the threat landscape enough that the remediation timeline its policy encouraged is insufficient. Project Zero says large language models (LLMs) let attackers locate and exploit weaknesses faster and increase the chances that many groups, including malicious actors, find the same flaw.

The group says a weakness reported now creates far more immediate danger for users than flaws reported last year did. Large language models shorten the time attackers need to reverse engineer and exploit patched flaws, so near-zero patch-to-exploitation time is a plausible model of how attackers behave, according to Project Zero.

The group said patch soak time no longer matters, since attackers can extract a bug’s details almost instantly from a patch and that removes any real secrecy. Patch gap is the time from a security fix’s release until a user installs the relevant update, per Project Zero.

Project Zero says vendors should narrow that window by accelerating rollout or applying other protections while remediation work continues. According to Project Zero, large language models have increased its vulnerability volume to many flaws a week in targets where the group once found only a few per year.

Under the “Full X” framework, Project Zero makes a flaw public X days after the vendor gets the report. The schedule starts with “Full 90” on Nov. 13, 2026, moves to “Full 60” on Feb. 2, 2027, and shifts to “Full 30” on Jan. 3, 2028.

Project Zero says the new disclosure policy aims for quicker remediation, a reduced patch gap, and automation from both Project Zero and vendors. While developing the new policy, Project Zero gathered feedback from the vendors that receive the most vulnerability reports.

Policy points that require vendor negotiation or agreement, including whether vendors fixed a bug or whether a grace period applies, are becoming unmanageable, the group says. The group plans to end nearly all exceptions from its disclosure rules, including the grace period and the variant policy.

The sole exceptions cover actively exploited vulnerabilities managed under Google’s application security vulnerability disclosure policy. If others have already published vulnerability details, Project Zero may release them at its discretion. It will also release vulnerability details if the vendor that received the report asks for disclosure.

Vendors warned that moving to the new policy immediately could leave users exposed to unfixed bugs while vendors upgrade remediation processes and technology. They said an overly aggressive disclosure policy could push them to rank newly reported flaws above other, less severe reported vulnerabilities. Vendors also warned that enforcing a policy some vendors cannot meet might lead them to abandon fixes before the deadline.

Project Zero says the benefits of delaying disclosure shrink because attackers are more likely to find a flaw during the embargo period, although short-term benefits remain. Project Zero chose a gradual schedule it describes as challenging but realistic, hoping vendors broadly improve remediation processes and infrastructure to fix all reported flaws quickly.

The group adds vulnerabilities to its tracker only when they show significant severity and demonstrated security impact, although its analysis is not always perfect. Project Zero says severity does not work well as a metric for disclosure decisions because it does not change how likely attackers are to find a flaw, only how it affects users.

Vendors decide what risks they will expose users to when they rank fixes by severity. Project Zero says it can be reasonable to accept the risk of low-severity bugs such as unexploitable null pointer crashes or minor information leaks.

The group applies high standards for reproducibility, readability, reachability and reported security impact to the flaws it reports. Vendors may handle low-quality reports differently.

Project Zero urges vendors to prepare for imminent disclosure of serious vulnerabilities, including responses that do not involve a patch, because defenders cannot count on finding a bug first. Project Zero still plans to track whether vendors fix flaws and to report when they do not, but fixed status will not serve as a decision point under the new policy.

By tracking fix status, Project Zero can explain how protection reaches users without issuing a binary judgment that carries consequences for vendors. Deciding whether vendors fixed a bug remains hard and unresolved, especially for code not shipped directly to users or not publicly available.

For each report, Project Zero often exchanges many messages with vendors to agree on whether vendors fixed a flaw. As vendors use non-patching remediation such as temporarily disabling code paths, Project Zero expects agreeing on whether vendors fixed a flaw to become more complex.

Project Zero says it will keep making rare policy exceptions. The group has sometimes made such exceptions, including Spectre and Meltdown.

On Oct. 1, 2026, Google said it would temporarily stop accepting product vulnerability submissions to its Open Source Software Vulnerability Rewards Program. Google pointed to a significant rise in automated submissions, most of them invalid, and said an update would arrive in Q1 2027.

Google said there was “a significant rise in automated submissions, the vast majority of which are not valid.” Open-source maintainers said that by mid-year, AI scanning tools had driven report volume to about four times historical levels.

Duplicate submissions and proofs of concept that fail to compile appeared alongside genuine high-severity findings. Reports that overlook a project’s threat model turned up in the surge.