High
Google, GTIG, AI Threat Tracker, AI agents, ARTEX, CAIRN, exploit_agent, Project Zero, patch diffing, n-day, LLMjacking

Google AI Threat Tracker: AI Agents Now Weaponize Patches
Google AI Threat Tracker: attackers use AI agents like ARTEX and CAIRN to turn patches into exploits, and Project Zero is cutting disclosure to 30 days.
Google's AI Threat Tracker says attackers now use AI agent frameworks to turn security patches into working exploits, in some cases with no human in the loop. Google Threat Intelligence Group (GTIG) named ARTEX, CAIRN, CyberStrikeAI and HexStrike AI among the tools in use, and Google Project Zero has started cutting its disclosure deadline from 90+30 days to 30 days, because AI is speeding up exploitation.
GTIG's report, published November 16 and covered by BleepingComputer and The Hacker News, says AI use by threat actors has moved past manual prompting. Financially motivated groups and suspected China-aligned espionage actors are running open-source, multi-agent penetration-testing frameworks against internet-exposed systems with low-complexity flaws that a large language model can write an exploit for.
The clearest case is a framework GTIG calls exploit_agent, found on an attacker's exposed staging server alongside stolen victim data, webshells and brute-forcing scripts. It watches software updates for popular web apps, works out what each patch fixed, and tries to exploit the flaw automatically. GTIG says it was used against French media, entertainment and education websites, and that this is the first time it has seen vulnerability discovery and exploitation fully automated with no human involved.
How are attackers using AI agents to exploit patches?
The technique is patch diffing at machine speed. When a vendor ships a fix, the old and new code sit side by side, along with release notes and any public research. GTIG says that finite set of material makes it efficient for an LLM pipeline to find the bug and draft a proof of concept. It cites the PaperCut NG/MF flaws in September as a case where this compressed the gap between patch and exploitation. The report also describes the noise this creates: after the PHP2Shell disclosure, hundreds of purported exploits appeared, most of them AI-generated, and defenders had to sort real ones from fakes.
Other findings in the report:
Finding | Detail |
|---|---|
Frameworks abused | ARTEX, CAIRN, CyberStrikeAI, HexStrike AI, PentAGI, Agent Zero |
Fully automated attacks | exploit_agent ran discovery and exploitation with no human in the loop |
Patch-to-exploit case | PaperCut NG/MF flaws, September 2026 |
Modified Claude Code | UNC6836 altered Claude Code settings on compromised machines to fetch the SOMBERMEME backdoor at launch, targeting crypto firms |
LLMjacking | Stolen LLM credentials resold through proxy services such as APIMart and JetNight |
Forecast | Unsophisticated attacks needing little or no human input become common within three months; sophisticated, fully autonomous attacks likely within six |
The FBI separately warned last week that several threat actors have used ARTEX against financial, government, technology, defense and healthcare organizations since August.
Why is Project Zero cutting its disclosure deadline to 30 days?
Because the embargo period is losing value. Project Zero told BleepingComputer that AI makes it more likely attackers find the same bug during the embargo and easier to reverse-engineer fixes once they ship. The change is staged so vendors can rebuild their remediation processes:
Date | Project Zero deadline |
|---|---|
Before Nov 13, 2026 | 90 days, plus 30-day patch adoption window |
From Nov 13, 2026 | 90 days, extension removed |
From Feb 2, 2027 | 60 days |
From Jan 3, 2028 | 30 days |
Details are published when the deadline expires, fixed or not. Exceptions apply for bugs exploited in the wild, bugs already public, and cases like Spectre and Meltdown where Google held details back longer. SecurityWeek notes vendors had warned that an immediate jump to 30 days could expose users while fixes were still in progress.
What should defenders do about AI-automated exploitation?
Assume the window between a patch and a working exploit is now days for anything internet-facing. Patch exposed web apps and edge devices first, using CISA's Known Exploited Vulnerabilities catalog and EPSS to rank, rather than waiting for a monthly cycle. Lock down LLM API keys and cloud credentials, since GTIG ties LLMjacking to the same ecosystem. Plan for change control too: at 30 days, a disclosure can go public while a fix is still waiting for change-board approval, so pre-approve an emergency path for internet-facing systems. Expect a flood of fake proofs of concept after big disclosures and validate before acting on any of them.
Our read
The tools in this report are the same class of agentic testing frameworks defenders use. What changed is cadence: attackers now test every patch the day it ships, while many organizations still verify their exposure once a year. Our analysis of 2025's KEV data found the median gap from disclosure to exploitation was 26 days, about 20% were exploited on or before disclosure day, and an annual pentest would have missed 67% of actively exploited vulnerabilities. For CVEs published in 2026, the median time to exploit is 5 days from disclosure to a KEV listing. Machine-speed attacks call for machine-speed verification. Agentic VAPT runs the same kind of continuous testing on your own systems, so you see a new exposure before an exploit_agent does. Related: the WorkOS breach involving ARTEX, LLMjacking and the PaperCut exploitation. Nexus Void runs continuous VAPT for web apps, APIs and cloud.
Reporting by BleepingComputer, The Hacker News and SecurityWeek; findings per Google Threat Intelligence Group.