Google and Anthropic push for faster patches as AI collapses bug discovery timelines
CSO Online ·

AI-assisted vulnerability discovery has made it untenable for software vendors to sit on privately reported flaws for months, and Google Project Zero is responding by shrinking its vulnerability disclosure deadline from 90 to 60 days, with plans to drop it to 30 days a year from now. Anthropic is following suit with an even more aggressive timeline. Project Zero is Google’s elite vulnerability research team that regularly reports vulnerabilities in widely used open-source and proprietary software, much of which is critical to internet infrastructure. Its 90-day deadline for vendors to produce patches before the team releases vulnerability details publicly was a major stake in the ground for the security research community when it was announced in 2014, and it has become the de facto standard for coordinated disclosure. The team’s announcement Friday that it will reduce this deadline to 60 days starting in February and 30 days starting in January 2028 is important because when Project Zero speaks, the security industry listens. And it did. On Monday, Anthropic announced it will adjust its disclosure policy from 90 days to 30 days for high- and critical-severity issues, and to 45 days for all other issues, starting in January. Other security companies and researchers will likely follow. The rationale is explicitly due to the offensive cyber capabilities of large language models: not only in discovering new vulnerabilities, but in reverse engineering patches and using them as blueprints for finding additional flaws in the same code . “LLMs have increased the speed at which attackers can find and exploit vulnerabilities, alongside the risk that multiple parties, including malicious ones, discover the same bug,” Project Zero researcher Natalie Silvanovich said in a blog post . “A vulnerability reported today presents a much higher immediate risk to users than one reported last year. In this new environment, it is crucial that vendors move towards remediating vulnerabilities faster.” Anthropic’s Frontier Red Team said essentially the same: “A 90-day window made sense when a vulnerability could be expected to stay private for months. We no longer believe it can, and these changes are intended to keep pace with progress in AI capabilities and strike the right balance for defenders.” ‘Soak time’ has become meaningless Since 2021, Project Zero’s policy has actually been “90+30” with 90 days to develop a patch, plus 30 days of grace while the patch rolled out to users. The team referred to these extra 30 days as “soak time.” But in a world where AI agents can analyze patches to determine flaw details in minutes and then write exploits for them, waiting for a patch to propagate before releasing vulnerability details only creates a false sense of secrecy. As a result, the Project Zero’s previous 30-day grace is gone effective immediately. In a report published this week, Google’s Threat Intelligence Group (GTIG) documented the first fully automated AI-powered vulnerability discovery-and-exploitation pipeline it has observed in the wild. Called exploit_agent, the framework is designed to analyze application code changes following security patches, identify vulnerabilities, and generate exploits without requiring an operator to direct each step. While Anthropic has different deadlines based on vulnerability severity, Project Zero decided to abandon severity as a disclosure metric because attackers are just as likely to discover low-risk flaws as they are to find the high-risk ones. “When vendors determine severity and use it to prioritize bug fixes, they are making decisions about what risks they are willing to expose their users to, and it is reasonable to accept the risk of low-severity bugs, like unexploitable null pointer crashes or minor info leaks in a lot of circumstances,” Project Zero’s Silvanovich said. “But this doesn’t change the availability of these vulnerabilities to attackers, and we want our policy to reflect this reality.” Google’s team will also no longer check whether a vendor’s patch fixes the issue when deciding whether to release details or not, because this back-and-forth can quickly become unmanageable, especially because the team now finds vulnerabilities using AI assistance at a much faster pace and because vendors might choose alternative mitigation strategies such as temporarily disabling a vulnerable feature or code path until a full patch can be developed. “We decided on a policy that encourages vendors to improve remediation time on a schedule that is challenging but realistic, with the hope that this will lead to vendors broadly improving their remediation processes and infrastructure in a way that eventually enables them to fix all reported vulnerabilities quickly,” Silvanovich said. The road ahead will be bumpy for users The shorter patch deadlines come at a time when open-source projects and software vendors are already drowning in vulnerability reports and must resort to either AI-assisted patching or hiring more developers, if they can afford it. As of last week, the number of CVEs issued this year exceeded 100,000, more than double last year’s count. Anthropic’s Frontier Red Team noted that “models can now help develop fixes” as one justification for the more aggressive disclosure deadlines, and said it will include “actionable guidance with each report, along with a patch or mitigation where possible,” to help maintainers confirm flaws and fix them. However, multiple studies and developers have reported in recent months that AI-generated patches are often incomplete, failing to address the root causes of a flaw, to eliminate all instances of the bug from the codebase, or to block all exploit variations. Security firm XOR built a benchmark that pitted frontier coding systems against vulnerabilities the company had discovered in a private software stack. No public fix existed for any of the flaws when the evaluation ran, so the systems had to work from the vulnerable code alone. The best models produced complete fixes for only about a third of the flaws: 35.3% for GPT-6 Astra and 28.6% for Claude Fable 5.1. Across all systems, just 11.9% of patch attempts earned full credit. According to an analysis by VulnCheck , seven months after launching its Mythos-powered Glasswing vulnerability-hunting program, Anthropic’s public ledger counts 37,966 findings, but only 217 of the 7,349 flaws made public so far have been fixed. Project Zero acknowledges the friction between the need to patch faster and devs being overloaded. The team solicited feedback from the vendors it reports to most and chose the phased schedule because an immediate change could push vendors to stop trying to meet the disclosure deadlines. “To respond to these new timelines, organizations should be better prepared to assess whether compensating controls can reduce the risk of a vulnerability when it’s disclosed,” Ramy Rahman, senior principal solutions engineer at ArmorCode, tells CSO. “With disclosures arriving sooner, and sometimes before a fix is available, teams need a process for evaluating exposure, applying available mitigations, and tracking remediation.” “Consider hardening systems at deployment time and removing software you don’t use,” he said. “Leaner, more deliberate deployments can reduce exposure and make vulnerability findings easier to manage.”