Google Threat Intelligence Group said Monday that, for the first time, it had found attackers using a language-model framework to discover and exploit software flaws with no human in the loop, a change it said can shrink the time between a public patch and a working exploit.

In a Nov. 16 report published with Mandiant, GTIG said it identified an exposed staging server running a multi-agent system it dubbed exploit_agent. The framework, the report said, “is set up to conduct continuous differential analysis of popular web application software. It compares code before and after patch announcements, analyzes findings to identify potential vulnerabilities, drafts exploits, and automatically executes them without human intervention.” The server also held “data exfiltrated from targeted victims”. Victim telemetry “suggests the operator was specifically interested in targeting French websites associated with the media and entertainment industries, as well as the education sector.” The report did not name those sites, count successful compromises, or identify who ran the server. Comments in a configuration file, published with GTIG’s translation, were written in Russian and told the operator to enter API keys for “auto-exploitation”, but the report did not attribute the framework to a country or a named group.

GTIG said attackers may find it more efficient to point models at already-patched “n-day” vulnerabilities than to hunt for unknown flaws, because a patch supplies a finite set of code and research to compare. In September, it said, overall vulnerability disclosures and exploits “nearly doubled between January and August 2026,” while zero-day exploits rose from a 2025 average of 8 a month to 11 a month in 2026, “before spiking to 22 in the month of August.” Once a vendor ships a fix, the clock starts on every unpatched system facing the internet. GTIG said security researchers and the cybersecurity company Huntress generated working exploits for two PaperCut print-server flaws within a day of disclosure. GreyNoise, the threat-intelligence firm, said in a Sept. 9 post that a likely Russian-speaking actor used artificial intelligence to develop, test and use exploits for PaperCut, and that where the attacker reached domain administrator, its fastest time was five minutes. In one case, GreyNoise said, the attacker went from initial access to full domain administrator in seven minutes against a U.S. high school.

A chart in the report, titled “Evolution of threat actor use of LLMs” and captioned Figure 2, places current activity at “STAGE 5”, described as “Multiple attack lifecycle stages conducted by LLM-driven tooling in-the-wild”, with “Minimal human-in-the-loop”. Under the heading “PROJECTED · NEXT 1–6 MONTHS”, it lists “STAGE 6” in the “Near-term (1–3 months)” as “Multi-agent-driven, unsophisticated attacks with little to no human interaction”, labeled “No human-in-the-loop”, and “STAGE 7” in the “Medium-term (3–6 months)” as “Sophisticated, fully-autonomous/agentic attacks”, labeled “Fully autonomous”. Those later stages are forecasts, not a pattern GTIG said it had already observed in the wild. The exploit_agent finding is narrower: automated vulnerability discovery and exploitation.

Separately, GTIG said adversaries are using open-source, multi-agent offensive frameworks, including ARTEX, CAIRN, CyberStrikeAI and PentAGI, “to execute multi-stage intrusions at greater operational velocity.” It said it had “independently observed” CAIRN and PentAGI “deployed in the wild”, and had “seen experimentation with CyberStrikeAI and HexStrike AI.” CAIRN was used in “intrusion activity against targets including in the aerospace sector in the US”, the report said. Between late August and early October, it tracked an actor using a customized platform built on PentAGI “to conduct systematic reconnaissance and vulnerability scanning against a wide range of organizations in Asia.” It named no victims in those cases. The tools, it said, “currently trade stealth for velocity.”

The FBI issued a separate advisory on ARTEX three days earlier. In FLASH-20261113-001, dated Nov. 13 and posted by the Internet Crime Complaint Center, the bureau said threat actors have “weaponized” the framework “to automate multiple phases of cyber intrusions at machine speed, from reconnaissance through data exfiltration.” “ARTEX significantly reduces the time required to chain attack techniques, with some intrusions progressing from initial targeting to privileged access in under three hours,” the advisory said. It cited “financially motivated attacks against South Korean financial institutions resulting in exfiltration of personal information from approximately 67,000 individuals,” and “a supply chain compromise affecting ConnectWise Cloud Backup customers that resulted in attempted sign-ins to more than 2,250 downstream tenants.” Those figures are the FBI’s, not GTIG’s. The Google report listed ARTEX among frameworks in use but did not attach those campaigns to it.

In a section on activity it associated with the Democratic People’s Republic of Korea, GTIG said a “suspected MIDNIGHT NEPTUNE” cluster had, in “recent intrusions targeting cryptocurrency organizations in North America and Europe,” altered `.claude/settings.json` files in internal repositories “so that when users started up Claude Code, it would automatically trigger the download of the SOMBERMEME backdoor.” Initial access, it said, came through fraudulent recruiter profiles whose imagery “contained SynthID watermarks, confirming use of LLMs to synthesize deceptive professional identities”, and “ClickFix” lures that prompted targets to run purported diagnostic scripts. To hide later movement, “suspected MIDNIGHT NEPTUNE” created GitHub personal access tokens “named after Microsoft Copilot”, the report said. GTIG framed the technique as abuse of a coding assistant’s configuration, not a vulnerability in Anthropic’s Claude Code product, and published no malware indicators, victim names or incident dates. As of Tuesday afternoon, Eastern time, no official Anthropic statement or quoted comment on the finding was found.

Financially motivated actors, GTIG said, continue to profit from LLMJacking: stealing language-model credentials and cloud resources from compromised companies and reselling access through underground “LLM Mixers”. “Throughout 2026, Mandiant responded to several LLMJacking incidents” in which actors consumed victims’ cloud or model quotas, the report said. “In some cases, single-day fraudulent charges exceeded $100,000 USD. In the most severe cases, cumulative resource abuse reached $1 million USD before remediation.” Mandiant and GTIG said they had seen the pattern in financial services, retail, transportation and technology, and did not name the companies.

Google, which published the research, also sells defenses against this class of threat. The report said the company uses “our autonomous Google AI Threat Defense architecture and BeyondZero”, and that it developed Google AI Threat Defense as an enterprise product. Google said it had disabled accounts and assets tied to some of the abuse it found. Those are the company’s accounts of its own products and response.