Anthropic Cuts Bug Disclosure to 30 Days as Patching Lags
Three days after Google's Project Zero scrapped its 90-plus-30 policy, Anthropic shortened its own window for high-severity flaws from 90 days to 30. Its public ledger shows 516 of 6,157 disclosed vulnerabilities patched.

Anthropic will give software vendors and open-source maintainers 30 days rather than 90 to fix high- and critical-severity vulnerabilities that its models find, under a policy change its Frontier Red Team published on 16 November 2026. All other severities move to 45 days. The new timings apply to vulnerabilities reported on or after 1 January 2027, and reports already in progress on that date keep their existing deadlines.
Two related clocks move with it. If a maintainer does not respond to an initial report, Anthropic will aim to escalate to an external vulnerability coordinator after 15 days instead of 30. And it will wait 30 days rather than 45 after a patch ships before publishing full technical details. The seven-day window for anything under active exploitation is unchanged, and the policy does not cover findings from its OSS Scanner service that no human has reviewed.
There is also a backlog to clear. Anthropic said it is notifying maintainers of high- and critical-severity flaws it reported more than 90 days ago, which now sit past the deadline in the current policy without being disclosed, that they have an additional 30 days before the reports go on its public dashboard.
Project Zero got there three days earlier
Google's Project Zero, which has used a "90 + 30" policy since 2021 (90 days to remediate plus 30 days of patch rollout soak time), announced on 13 November 2026 that it is switching to a fixed clock it calls Full X. Vulnerabilities are now published 90 days after they are reported to the vendor regardless of whether a fix exists, dropping to 60 days on 2 February 2027 and 30 days on 3 January 2028. Most special cases go, including the grace period and the variant policy.
Natalie Silvanovich of Project Zero wrote that the advent of AI-assisted vulnerability discovery and exploitation has changed the threat environment "such that the vulnerability remediation timeline encouraged by our disclosure policy is clearly insufficient". The post also records the volume shift behind the decision: "We are finding several vulnerabilities a week in targets where we once found a few vulnerabilities a year."
The two policies differ on one point of principle. Anthropic keeps severity tiers, giving low-severity bugs 45 days. Project Zero explicitly rejects that: "A vulnerability's severity does not affect the likelihood that it will be discovered by attackers, only its impact on users." Google's reasoning is that soak time has stopped working, because attackers can read a patch and build an exploit from it almost immediately.
The ledger Anthropic keeps on itself
Anthropic publishes a running count of what its programme has produced, and the numbers explain why the deadline fight is sharp. As of 2 October 2026, the dashboard records 6,157 vulnerabilities disclosed across 591 open-source projects, of which 516 have been patched to Anthropic's knowledge. That is about 8 per cent. Maintainers had acknowledged 5,103 of the reports. Across all findings, 584 identifiers had been issued: 219 CVE records and 365 GitHub Security Advisories.
The funnel behind those figures is worth reading too. The same dashboard counts 29,439 candidate findings, of which 6,123 went to external security research firms for review and 5,674 were confirmed valid, a true positive rate of 92.7 per cent. Anthropic reported 4,824 findings direct to maintainers without that independent check, and notes those may contain false positives.
The honest reading is that the constraint has moved. Discovery is no longer the scarce resource; triage and repair are, and both are still done largely by people. Anthropic's own ledger shows a programme generating findings roughly twelve times faster than the ecosystem patches them, and the policy change tightens the deadline on the slow half of that equation rather than the fast one. Anthropic argues the opposite case plainly: a bug one model can find is increasingly likely to be found by another, so a 90-day embargo now mostly protects whoever rediscovers it first.
Whether 30 days is achievable depends who you ask
Katie Moussouris, chief executive of Luta Security and one of the people who built modern disclosure practice, told Dark Reading on 16 November that the gap is the problem. "A deadline doesn't magically manifest a fix," she said, adding that "researchers and AI companies are burying maintainers in bug reports and not doing enough to help fix them". Dustin Childs, head of threat awareness at the Zero Day Initiative, raised a different risk in the same report: "I fear it will lead to patches that don't fully address the vulnerability. We see that happen at 120 days."
The strongest counter-evidence comes from Google. A spokesperson told Dark Reading that the company consulted the vendors Project Zero reports to most, including open-source projects, and that "every vulnerability that Project Zero reported against a volunteer-maintained project in 2026 was fixed in 30 days, except for one, which was fixed in 31 days". On that record, volunteer projects need to change least. The caveat is selection: Project Zero reports a small number of high-impact bugs to a narrow set of targets, which is not the position of a maintainer receiving an automated batch.
Alex Rice, co-founder and chief technology officer of HackerOne, told Dark Reading that specifics will keep varying by programme and severity, "but the direction isn't in question. Organizations should plan for 30 days as the baseline expectation that will shrink even further." What nobody has published is the figure that would settle the argument: how often a shortened deadline produced a fix that would not otherwise have shipped, as against how often it produced a public advisory with no patch behind it. Anthropic's first cohort under the new rules will not reach its deadline until February 2027.
Anthropic says it will keep sending actionable guidance and, where it can, a patch or mitigation with each report, and will agree a sustainable pace with a project before sending a large volume of findings. It also points maintainers at free tooling, including six months of its top subscription tier for eligible open-source contributors. Whether that offsets a two-thirds cut in the deadline is the question the next quarter answers.
What happens next?
- Anthropic's shortened windows apply to vulnerabilities reported on or after 1 January 2027, with the first 30-day deadlines falling due in early February.
- Project Zero moves to a 60-day fixed clock on 2 February 2027 and a 30-day clock on 3 January 2028.
- Maintainers of high- and critical-severity flaws Anthropic reported more than 90 days ago have 30 days before those reports are published on its dashboard.
- Watch whether other large vulnerability research programmes, including vendor-agnostic ones that currently use 120 days, follow the 30-day baseline.
Related topics
Sources & references
- 01Updating our coordinated vulnerability disclosure policy timelines — AnthropicprimaryFrontier Red Team post, 16 November 2026.
- 02Policy and Disclosure: 2026 Edition — Google Project ZeroprimaryPublished 13 November 2026; source of the Full X schedule and the Silvanovich quotes.
- 03Coordinated vulnerability disclosure dashboard — AnthropicdataCounts as of 2 October 2026: 6,157 disclosed, 591 projects, 516 patched, 584 identifiers.
- 04Google Project Zero Slashes Bug Disclosure Deadlines, Splitting Experts — Dark ReadingnewsAlexander Culafi, 16 November 2026; source of the Moussouris, Childs, Rice and Google spokesperson quotes.
- 05Google Project Zero to Cut Disclosure Deadline to 30 Days by 2028 — Infosecurity MagazinenewsIndependent account of the 13 November policy change.
More from Security
France Names an IRGC Cyber Unit Behind 700 Fridge Attacks
France attributed the October attacks on refrigeration controllers at more than 700 retail sites to an offensive cyber unit of Iran's Islamic Revolutionary Guard Corps on 9 November 2026, summoned the Iranian ambassador and said it will propose further European sanctions.
Five US Agencies Warn on Exposed Refrigeration Controllers
CISA, the FBI, HHS, USDA and the Department of Transportation told owners of industrial refrigeration and cold chain monitoring systems to take controllers off the public internet. The one quantified sign of progress, a 53 per cent fall in exposed devices, came from a scanning company rather than the government.
Binance Saw $3.64bn of Net Outflows on Its First Day Back
Binance says it has replaced 95 per cent of the stolen tokens with its own money, having set aside $6.34 billion. The FBI has attributed the largest cryptocurrency theft on record to North Korea, and about $3 billion of the proceeds was frozen on-chain.
Stolen GitHub Tokens Put Malware in a LiteLLM Release
A compromised release pipeline put a poisoned LiteLLM build on PyPI and Docker for under three hours on 21 October 2026. It harvested cloud secrets, AI provider API keys and SSH keys, and Wiz says the attackers were spending stolen AWS credentials about four and a half hours later.



