US Edition
Your source for latest news
TechnologyCybersecurity

Google Says Attackers Built an AI Tool That Turns Patches Into Exploits With No Human Involved

Google's threat-intelligence group says it has found the first case of a fully autonomous AI framework that reads software patches, writes working exploits and launches attacks without a person in the loop, part of a broader surge in AI-driven intrusions it expects to accelerate within months.

PT
By PressTemps Technology DeskPublished Today, 12:39 ET · 5 min read
Google Says Attackers Built an AI Tool That Turns Patches Into Exploits With No Human Involved
File photo. A generic illustration of malicious code — not an image of the exploit_agent framework itself, which has not been publicly depicted. Photo: Santeri Viinamäki / Wikimedia Commons, CC BY-SA 4.0
What to know
Google's Threat Intelligence Group says it found the first case of an AI framework, "exploit_agent," that autonomously turns newly patched software flaws into working exploits with no human involved.
Disclosed and exploited vulnerabilities nearly doubled from January to August 2026; monthly zero-day exploitation rose from an average of 8 in 2025 to 22 in August 2026 alone.
Other autonomous attack frameworks named in the report, including CAIRN and PentAGI, have already been used against US aerospace targets and organizations across Asia.
Google forecasts that largely autonomous, low-sophistication AI attacks will be common within three months and fully autonomous intrusions likely within six.

Google's threat-intelligence arm says it has found the first documented case of attackers using artificial intelligence to turn a software patch into a working exploit with no person involved in the process.

In a report published on 16 November, the Google Threat Intelligence Group (GTIG) said it discovered an exposed staging server running a framework its operators called "exploit_agent." The tool, described in full in GTIG's quarterly AI Threat Tracker, continuously compares the before-and-after code of popular web applications whenever a security patch is released, works out what flaw was fixed, drafts an exploit for the unpatched version, and runs it against targets automatically.

An exploit with no human in the loop

GTIG researchers Aurora Blum, Kelli Vanderlee, Jess Xia and David Mainor wrote that the discovery marks the first time the company has observed threat actors fully automating both vulnerability discovery and exploitation in this way. The framework stitches together several large language models, both frontier commercial systems and smaller local ones, and was found alongside stolen victim data, reconnaissance notes, brute-forcing scripts and tools for planting or removing webshells. Its configuration file carried comments written in Russian, including an instruction, translated by Google, to "enter the keys yourself, right here, not in the chat," referring to the API keys needed to run the automated exploitation.

Telemetry on the server suggested the operator had been probing French media, entertainment and education websites, though Google stopped short of naming a country or group behind the tool, and VentureBeat reported that the report does not establish the framework actually caused the compromises found on the same server.

The numbers behind the warning

GTIG's report sets exploit_agent against a broader rise in automated attack activity. Disclosed and exploited vulnerabilities nearly doubled between January and August 2026, the company said. Zero-day exploitation, rarer and harder to automate, rose more modestly: an average of eight a month in 2025, climbing to 11 a month across 2026 before spiking to 22 in August alone.

Two of those August zero-days hit the print-management software PaperCut, which Google and outside researchers say were chained together for unauthenticated remote code execution; both flaws were added to the Cybersecurity and Infrastructure Security Agency's Known Exploited Vulnerabilities catalog after GreyNoise detected a campaign against hundreds of organizations within four days of disclosure, with some intrusions reaching domain-administrator access within minutes of initial entry. A separate pair of PHP flaws, patched on 21 October, triggered hundreds of claimed proof-of-concept exploits within days, most of them non-functional but apparently machine-generated, according to the GTIG report and Security Affairs' account of it.

A crowded field of offensive AI agents

Exploit_agent is not the only such tool GTIG is tracking. The report names several open-source, multi-agent offensive frameworks it says attackers have already put to use, including CAIRN, which Google observed deployed against United States aerospace targets, and PentAGI, used by one actor against organizations across Asia between late August and early October. Others, such as CyberStrikeAI and HexStrike AI, were seen being tested rather than deployed. GTIG also said a China-linked group appeared to be building its own automated attack framework, likely for espionage.

The report separately describes a case in which a group it links to North Korea, tracked as part of the "Midnight Neptune" cluster, modified a developer's Claude Code configuration file so that launching the coding assistant silently installed a backdoor, and notes that criminals are also renting out stolen access to AI models on underground marketplaces, with Mandiant tracing single-day fraudulent charges exceeding $100,000 in some incidents.

Who is exposed, and how defenders are responding

The immediate risk GTIG describes falls on organizations running widely used, internet-facing web software, since exploit_agent and its peers are built to scan for freshly patched vulnerabilities at scale rather than to pursue a single target. GTIG's own assessment is that these frameworks currently "trade stealth for velocity," meaning they leave detectable traces even as they shrink the window between a patch's release and a working attack. The firm says it has already shipped behavioral detection rules for one of the publicly known frameworks, ARTEX, through Google SecOps, and that Mandiant has published separate hardening guidance.

Outside security researchers who reviewed the findings framed the shift as an acceleration of an existing problem rather than something categorically new.

"Attackers didn't invent a new kind of bug. They automated the tedious part: reading the patch, finding what it fixed, and writing the exploit. A skilled person might spend days on that. Now an agent does it while its operator sleeps." — Merritt Baer, former deputy chief information security officer at Amazon Web Services, quoted by VentureBeat

Baer argued that the defensive response does not need to be exotic: organizations should know what systems they expose to the internet, patch faster than in previous years, and monitor for signs of active exploitation rather than relying on vulnerability scanning alone. BleepingComputer, which also reviewed the report, noted that the compressed patch-to-exploit timeline is the main practical consequence for security teams, since it narrows the period in which a newly released fix can be applied before an automated attacker finds a way around it.

What Google expects next

GTIG's forecast is explicitly time-bound. The group said it considers it "highly likely" that unsophisticated, largely autonomous attacks requiring little human interaction will become common within three months, and that more sophisticated, fully autonomous intrusions are likely to emerge within six. It added that stealthier, long-running espionage operations are less likely to adopt noisy automated tools quickly, while profit-driven cybercrime groups, which care less about getting caught, are expected to keep adopting them regardless of success rate, because the frameworks let a single operator run many more attempts than before.

Google said it is feeding the findings back into its own defensive products, including the Gemini-based threat-detection system AI Threat Defense and the automated vulnerability-discovery and patching tools Big Sleep and CodeMender, and that newer models such as Gemini 4 Carbon have been stress-tested against misuse by outside groups including Gray Swan and Trajectory Labs before release. The company has not said when, or whether, it will identify who built or ran exploit_agent.

More on this story

All Technology