Cybersecurity

Google cuts bug-fix deadlines as AI starts turning security patches into attacks

Adrian Kessler
Add us on Google

Every security update tells attackers where the hole was, and some of them now have software that reads it for them. Google‘s threat intelligence team found a criminal setup that compared web applications before and after a security fix, worked out what had been repaired, wrote a working attack and launched it, with no person involved at any step. Google’s own bug hunters at Project Zero are responding by giving software makers less time to fix the flaws they report.

For anyone who taps “remind me later” on an update, the safe delay has shrunk to almost nothing. Project Zero now treats an attack arriving the moment a patch ships as a reasonable model of how criminals behave. The old idea of a quiet period after a fix, while people get around to installing it, no longer holds when a machine can reverse-engineer the patch faster than most users open their settings.

Since 2021, Project Zero has given vendors 90 days to fix a reported flaw plus an extra 30 days of “soak time” for users to install the patch. That extra month is gone. Flaw details are now published 90 days after the report, fixed or not, and the deadline will tighten twice more. The team is also dropping most special cases, including grace periods. It says AI has raised its own output to several vulnerabilities a week in products where it once found a few a year, too many to negotiate one by one.

Anthropic, whose models are among those now finding such bugs, is moving in the same direction. Its Frontier Red Team will publish high- and critical-severity flaws 30 days after reporting them instead of 90, and other flaws after 45. If a maintainer ignores a first report, it will call in an outside coordinator after 15 days rather than 30. “A 90-day window made sense when a vulnerability could be expected to stay private for months. We no longer believe it can,” the company wrote. A bug one model can find is increasingly likely to be found by another, it added.

The attack that alarmed Google sits in its latest AI Threat Tracker report. Investigators found an exposed staging server running a multi-agent framework called exploit_agent, built with Russian-language tooling and hooked up to both commercial and locally run AI models. It continuously compared versions of popular web software, and the same server held data stolen from victims. According to The Register, telemetry showed the operator going after French websites in media, entertainment and education. Google also lists other AI attack frameworks in criminal use, including ARTEX, CAIRN, HexStrike AI and PentAGI, used by profit-driven gangs and suspected espionage groups alike.

Shorter deadlines do not make anyone patch faster on their own. Vendors told Google that jumping straight to 30 days could hand attackers unfixed flaws while companies are still rebuilding their repair processes, and could push teams to rush Project Zero reports ahead of more serious bugs from other researchers. That is why Google is phasing the change in. Anthropic makes exceptions for critical infrastructure and for flaws that need a technical standard to change, and its policy does not cover AI findings no human has reviewed. Google, for its part, published the warning alongside a pitch for its own paid defense product, Google AI Threat Defense.

Neither policy is regional. Both apply to any company whose software these teams investigate, from phone makers to open-source projects run by volunteers, so the pressure lands on products used in every market. The slowest link remains the one Project Zero calls the patch gap: the time between a fix being released and the update actually reaching your device, which can depend on phone makers, carriers and app developers as much as on you.

Project Zero’s 90-day rule took effect on November 13, 2026. On February 2, 2027, the window drops to 60 days, and on January 3, 2028, to 30. Anthropic’s 30- and 45-day deadlines apply to reports filed on or after January 1, 2027. Maintainers already past their old 90-day deadline have been given 30 more days. Both teams keep a separate, faster track for flaws already under active attack, seven days in Anthropic’s case.

By 2028, the update you leave for tomorrow may already have an exploit written for it.

Tags: , , , , ,

Add us on Google

Discussion

There are 0 comments.