Google Threat Intelligence Group says it has seen attackers use AI frameworks to fully automate vulnerability discovery and exploitation, with no human in the loop, for the first time. In the quarterly AI Threat Tracker that GTIG and Mandiant published on November 16, the group said it found a multi-agent framework called exploit_agent on an exposed staging server.
The framework connects to multiple frontier and local models and is set up for continuous differential analysis of popular web application software, according to the report. It compares code before and after patch announcements, identifies potential vulnerabilities, drafts exploits and “automatically executes them without human intervention.” The server also held data taken from victims and scripts for internal reconnaissance, brute-forcing, installing webshells and removing other attackers’ webshells.
GTIG said victim telemetry on the server suggests the operator was interested in French websites in media and entertainment and in the education sector. Comments in an environment file that configures the framework’s model API keys are written in Russian; GTIG’s translation of one reads “LLM API keys for auto-exploitation (exploit_agent pilot).” A figure in the report shows the framework inspecting access-control code in version 2.19.25 of the Spectra WordPress plugin. GTIG does not name the operator or attribute the activity to a country.
GTIG’s figures show zero-days exploited rising from a 2025 average of 8 a month to 11 a month in 2026, with 22 in August alone. Overall disclosures and exploits nearly doubled between January and August 2026, the group said.
In late August, attackers exploited two PaperCut flaws as zero-days, and security researchers and Huntress used AI models to build working exploits within a day of disclosure, according to the report. GreyNoise reported a campaign that took place four days after disclosure, in which an attacker likely used AI models to build and test exploits, assemble a target list and attack hundreds of organizations. After the two critical PHP flaws were patched on October 21, hundreds of nonworking proof-of-concept exploits, likely generated with AI help, were published, GTIG said.
GTIG also observed attackers running open-source agentic penetration-testing frameworks during live intrusions. Besides ARTEX, these included CAIRN, used against targets including in the US aerospace sector, and a customized version of PentAGI, used for reconnaissance and vulnerability scanning against organizations in Asia from late August to early October. GTIG said it found evidence that a PRC-nexus group was developing its own automated attack framework, and that such frameworks “currently trade stealth for velocity,” leaving forensic traffic and code-level traces. Google Security Operations has deployed detection rules for ARTEX, and Mandiant has released a hardening guide for defending against AI-enabled offensive tools, the report said.
Throughout 2026, Mandiant responded to several cases of what GTIG calls LLMJacking, in which attackers broke into cloud environments to consume victims’ compute and AI model quotas. Single-day fraudulent charges exceeded $100,000 in some cases, and in the worst cases cumulative abuse reached $1 million before remediation. One attacker exploited CVE-2026-40217, a sandbox escape in LiteLLM that turns limited inference keys into full administrative credentials. Stolen access is resold through “LLM Mixers,” storefronts where GTIG found prices from $0.10 to $25 per million tokens.
In a section on North Korean activity, GTIG described intrusions against cryptocurrency organizations in North America and Europe by a suspected MIDNIGHT NEPTUNE cluster. The attackers approached targets through fake recruiter profiles whose images carried SynthID watermarks, which GTIG said confirmed the use of AI. After gaining access, they modified .claude/settings.json files in internal code repositories so that starting Claude Code would download a backdoor GTIG calls SOMBERMEME.
Sources: Google Threat Intelligence Group
–
By the Control Plane Editorial Team