AI Turns Security Patches Into Working Exploits in 30 Minutes. The 90-Day Window Is Over.
Agentic AISecurity analyst Himanshu Anand used a language model to turn a published React security patch into a working exploit in 30 minutes. Experienced reverse engineers used to need days to do the same thing. That gap — days to 30 minutes — is not a marginal improvement. It breaks the model the entire vulnerability disclosure industry is built on.
How the 90-Day Window Works (and Why It’s Failing)
Google’s Project Zero popularized the 90-day coordinated disclosure timeline: a researcher finds a vulnerability, notifies the vendor privately, the vendor has 90 days to ship a patch before the researcher publishes details publicly. The model rests on four assumptions: a single discoverer, sequential discovery timelines, a vendor head start that can be preserved, and a meaningful lag between patch release and exploit availability.
Anand’s research argues all four assumptions have collapsed under AI. Multiple researchers now find identical flaws simultaneously using the same LLM tools — he documented receiving the eleventh identical report of a critical purchase vulnerability within six weeks of the first. Patches, once released, are immediately analyzable for what they changed and what that change implies about the underlying flaw. The time from “patch released” to “working exploit” has, in the React case, become 30 minutes.
Documented Failures of the Disclosure Model
The evidence Anand presents is not theoretical. The “Copy Fail” Linux kernel vulnerability — discovered via a one-hour AI-assisted scan — had been present since 2017 and provided root access across Linux distributions. Iranian threat actors exploited it within days of public disclosure.
The “Dirty Frag” vulnerability illustrates embargo failure: a five-day coordinated embargo was broken within hours when independent researchers published their own findings. In a world where multiple parties can find the same flaw simultaneously with LLM assistance, embargos cannot be negotiated with all of them.
The Pattern
Over the past six months, Stack Archive has tracked a consistent theme across the agentic AI security landscape: AI capabilities doubling the attack surface, agentic systems creating new perimeter vulnerabilities, and DeepMind documenting autonomous agent hijacking at scale. The Anand research adds a specific, quantified data point: the offensive speed advantage AI provides is now measurable in minutes, not quarters.
The So What
Anand’s recommendations are direct: vendors should treat critical vulnerabilities as P0 emergencies requiring immediate patching rather than 90-day timelines. Researchers should pursue shorter disclosure windows. Administrators should apply patches within hours, not weeks. Development teams should use language models defensively — scanning their own dependency patches before attackers do — accepting that “attackers are winning that race” as the current baseline.
The agentic AI engineering discipline question this raises is architectural: if any published security patch can become a working exploit in 30 minutes using widely available LLM tools, the downstream assumption — that patch deployment timelines measured in days are acceptable — no longer holds. Security posture has to be rebuilt around the new attack tempo, not the old one.
Photo by Nikita Belokhonov on Pexels
Source: The Decoder
Content created with AI assistance and reviewed for accuracy.
Join the conversation
Stack Insiders is our free community for readers who want to go deeper — share resources, ask questions, and connect with others across every vertical we cover.
Join Stack Insiders →