Vulnerabilities: Google Reduces Disclosure Timeline to 30 Days by 2028

Google Project Zero will reduce the timeline for publishing details of a vulnerability to 60 days starting in February, and then to 30 days by January 2028. Anthropic also announces shortened deadlines, adjusted according to severity. Publishers will have to navigate these rules as AI assistance does not guarantee a complete fix.
Automated Fixes Often Remain Incomplete
The threshold of 100,000 CVEs published this year was crossed last week, more than double the total from the previous year. Open-source projects and publishers are already facing an influx of reports, which they are addressing with AI assistance or by hiring when finances allow. However, several studies and developers have reported incomplete automated repairs in recent months: the initial cause may persist, some instances may escape the fix, or exploit variants may still be possible.
The cybersecurity company XOR confronted advanced programming systems with vulnerabilities discovered in a private software suite. No fix was public at the time of evaluation: only the vulnerable code was provided to them. According to XOR, GPT-6 Astra fully fixed 35.3% of the flaws, compared to 28.6% for Claude Fable 5.1. Across all systems, 11.9% of attempts received full points.
VulnCheck also notes a gap between discoveries and repairs in Anthropic's public registry. Seven months after the launch of Glasswing, its vulnerability research program using Mythos, this registry reportedly contains 37,966 discoveries. Among the 7,349 vulnerabilities already public, 217 have been fixed.
Google Removes Conditions for Publication
Since 2021, publishers had 90 days at Project Zero to develop a fix, followed by 30 days dedicated to disseminating it to users. This grace period will disappear immediately. Google's research team, which regularly reports vulnerabilities in widely used open or proprietary software, particularly addresses critical components of internet infrastructure. Its 90-day deadline, introduced in 2014, had become a benchmark practice for coordinated disclosure.
The timeline announced on Friday staggers the reduction of the main deadline between February and January 2028. Project Zero acknowledges the overload on developers and has gathered feedback from the publishers to whom it sends the most reports. The team explains that it preferred this transition to an immediate change that could lead them to abandon their efforts to meet deadlines.
Severity will no longer determine the publication date. Natalie Silvanovich, a researcher at Project Zero, believes it is legitimate for a publisher to sometimes accept the risk of minor information leaks or non-exploitable crashes due to a null pointer. However, she explains that this prioritization does not make these bugs less accessible to attackers: their discovery remains possible regardless of the risk level.
Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisissez votre rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
Project Zero will also not condition its decision to publish on verifying the effectiveness of the fix. The team believes that these exchanges can become unmanageable, especially as it finds flaws much faster with the help of AI. It also emphasizes that the publisher can choose to temporarily neutralize a vulnerable function or portion of code rather than immediately deliver a complete fix.
Attackers Can Exploit Published Fixes
The justification put forth by Google and Anthropic concerns both the discovery of bugs and the examination of fixes by large language models. These can be used to understand a flaw and then search for others in the same code. Natalie Silvanovich asserts that attackers find and exploit vulnerabilities more quickly with these tools, while several parties, including malicious ones, are at greater risk of discovering the same defect. She thus estimates the immediate danger of a report to be significantly higher than last year.
According to Project Zero, agents can deduce the details of a flaw from a fix in a matter of minutes and then produce code to exploit it. Google's threat intelligence group, GTIG, presented this week exploit_agent, the first fully automated AI discovery and exploitation device it claims to have observed in real-world conditions. This device is designed to examine changes made by security updates, identify flaws, and generate exploits without an operator guiding each step.
Anthropic Promises Assistance with Each Report
Unlike Google, Anthropic will maintain a distinction based on severity. The company announced on Monday that starting in January, high or critical issues would be subject to a 30-day deadline, instead of the previous 90 days. For others, it plans 45 days. Other researchers and security companies may also revise their practices.
Frontier Red, Anthropic's team, cites the capability of models to contribute to repairs as justification for this acceleration. It promises concrete recommendations accompanying each report to facilitate the confirmation and then the correction of the issue, as well as a fix or mitigation when possible. At Google, Natalie Silvanovich describes a demanding but realistic timeline, which she hopes will push publishers to improve their procedures and repair infrastructures.
ArmorCode Recommends Preparing Temporary Protections
Ramy Rahman, a principal solutions engineer at ArmorCode, emphasizes that disclosures may occur while no fix is yet available. He recommends that organizations prepare to assess compensatory protections that may reduce risk. Teams should, according to him, have a procedure in place to measure their exposure, apply available mitigations, and track the repair.
He also advocates for hardening systems from the moment of deployment and removing unused software. Streamlined and better-controlled installations can, he believes, limit exposure while facilitating the handling of discovered flaws.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.