Anthropic Cuts Vulnerability Disclosure Window from 90 to 30 Days Because of AI
Anthropic is shortening the window for fixing discovered vulnerabilities: serious and critical flaws now have 30 days instead of 90, and all others get 45. The company argues that AI has sped up bug hunting, so a secret flaw can no longer be kept hidden for months. The rules apply to reports submitted from January 1, 2027. Maintainers should therefore set up fast report intake and patch release processes in advance.
AI-processed from TNW; edited by Hamidun News
When a researcher finds a vulnerability in someone else's software, they first notify the authors and give them time to fix it, and only then publish the details. For years, the standard was 90 days. Now Anthropic says this deadline no longer matches reality: AI models find bugs faster than before, and keeping them secret for three months is no longer safe. For those who maintain open-source libraries or ship commercial software, this is a direct cut to their response time.
What Exactly Is Changing
The changes were announced by Anthropic's Frontier Red Team. Serious and critical vulnerabilities will be published 30 days after a report, instead of 90. For all other issues, the deadline will be 45 days. The new rules apply to reports sent on January 1, 2027 and later. Reports already in progress will keep their previous deadlines.
Two more points are changing. If a maintainer has not replied to the first email, Anthropic will contact an external coordinator after 15 days, not 30. And after a patch is released, the company will wait 30 days instead of 45 before publishing the full technical description of the bug.
The seven-day window for vulnerabilities already being exploited in real attacks stays unchanged. Here, the clock was already counted in days.
Why 90 Days Stopped Working
Anthropic explains it this way: "The 90-day window made sense when you could expect a vulnerability to stay open for months. We no longer believe that." According to the company, the old rules rested on two assumptions. Patches are released at human speed, and two different people rarely find the same bug. With AI, the second assumption weakens: if one model has found a flaw, another, the company reasons, may find it too. In that case, secrecy for months protects little, while extra weeks without a patch give attackers a chance.
There is also a backlog. Some serious vulnerabilities that Anthropic reported more than 90 days ago remain unpublished, which makes them formally overdue. The company is giving these maintainers another 30 days. After that, the reports will appear on its dashboard under the new policy, in effect from January 1, 2027.
Exceptions and Help for Maintainers
Anthropic acknowledges that not everything can be fixed within 30 to 45 days. This applies, for example, to critical infrastructure software or to errors in technical standards, where changes take a long time to agree on. In such cases, the company promises to negotiate an individual timeline with the vendor.
The policy does not cover AI findings that no human has verified. It refers to raw output from automated scanning that nobody has confirmed.
The company admits that short deadlines will add pressure on maintainers, especially in small projects run by volunteers. To ease this, it will attach remediation guidance to each report and, where possible, a ready-made patch.
Not Just Anthropic
The trend extends beyond one company. Google's Project Zero is also shortening its disclosure window. Starting February 2, 2027, it will be 60 days, and from January 3, 2028, it will drop to 30. Two prominent vulnerability research groups are moving in the same direction.
What to Do in Practice
Teams that maintain public code should check three things. First, whether there is a security contact (for example, a security.md file or an email address) that someone actually reads: 15 days of silence now leads to escalation faster. Second, whether you can ship a patch within a couple of weeks, including tests and a release. Third, who on the project is responsible for coordinating with researchers.
For product companies, the conclusion is similar. Open-source dependencies that your stack relies on will get public vulnerability descriptions sooner. You will need to update them more quickly, and a "we'll update it next quarter" process becomes risky. It remains an open question whether small projects will have the resources to meet these deadlines. Anthropic's current answer is a promise of help with patches, but not a guarantee.
- Nov 21, 2026Калифорния засудила OpenAI за ИИ-агентов, взломавших Hugging Face: штраф $2500 за нарушение
- Nov 21, 2026Anthropic Heads to IPO at $1.4 Trillion, but Seven Founders Keep 50.1% of the Votes
- Nov 21, 2026A Ghanaian medical AI with its safeguards removed designed 1,020 proteins and 5,000 toxin molecules
- Nov 21, 2026A game without AI outranked 2,782 projects in Epic Games' AI jam popularity ranking
Key Questions Answered
What Exactly Is Changing?
The changes were announced by Anthropic's Frontier Red Team. Serious and critical vulnerabilities will be published 30 days after a report, instead of 90. For all other issues, the deadline will be 45 days. The new rules apply to reports sent on January 1, 2027 and later. Reports already in progress will keep their previous deadlines. Two more points are changing. If a maintainer has not replied to the first email, Anthropic will contact an external coordinator after 15 days, not 30.
Why 90 Days Stopped Working?
Anthropic explains it this way: "The 90-day window made sense when you could expect a vulnerability to stay open for months. We no longer believe that." According to the company, the old rules rested on two assumptions. Patches are released at human speed, and two different people rarely find the same bug. With AI, the second assumption weakens: if one model has found a flaw, another, the company reasons, may find it too.
What are the key details on Exceptions and Help for Maintainers?
Anthropic acknowledges that not everything can be fixed within 30 to 45 days. This applies, for example, to critical infrastructure software or to errors in technical standards, where changes take a long time to agree on. In such cases, the company promises to negotiate an individual timeline with the vendor. The policy does not cover AI findings that no human has verified. It refers to raw output from automated scanning that nobody has confirmed.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.