The short answer
- On November 16, 2026, Google Threat Intelligence Group (GTIG) said it identified threat actors using large language model (LLM) frameworks “to fully automate vulnerability discovery and exploitation without a human in the loop for the first time.”
- Google says GTIG found exploit_agent on an exposed staging server. It compared code before and after patch announcements, drafted exploits, and ran them without human intervention, Google said, and victim telemetry pointed to French media, entertainment, and education websites.
- According to Google, GTIG observed CAIRN used against targets including in the US aerospace sector, and tracked a customized PentAGI platform used for reconnaissance and vulnerability scanning against organizations in Asia, alongside ARTEX and CyberStrikeAI.
- Google said overall disclosures and exploits nearly doubled between January and August 2026, while zero-day exploits rose from a 2025 average of eight a month to 11 a month in 2026, then reached 22 in August 2026.
- In Google's North Korea section, a suspected MIDNIGHT NEPTUNE cluster, according to Google, modified .claude/settings.json in internal repositories so that starting Claude Code downloaded the SOMBERMEME backdoor.
- Google says these frameworks shrink patch windows and currently trade stealth for velocity, and that Mandiant Consulting released a hardening guide. Google Security Operations also deployed rulepacks to detect ARTEX.
Google Threat Intelligence Group (GTIG) said on November 16, 2026 that it identified threat actors using large language model (LLM) frameworks “to fully automate vulnerability discovery and exploitation without a human in the loop for the first time.” In a Google Cloud Blog post of the same date, titled “GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold” and written by Aurora Blum, Kelli Vanderlee, Jess Xia, and David Mainor, GTIG and Mandiant described an exposed staging server hosting a multi-agent framework Google calls exploit_agent.
The Register, in a November 16, 2026 report by Jessica Lyons, also covered the post and described it as Google's fourth-quarter 2026 AI threat tracker.
#What did Google say exploit_agent did?
Google says GTIG found an exposed staging server hosting a multi-agent vulnerability research and weaponization framework, which it calls exploit_agent, that integrates with multiple frontier and local LLMs. Google wrote: “The framework is set up to conduct continuous differential analysis of popular web application software.” It also wrote: “It compares code before and after patch announcements, analyzes findings to identify potential vulnerabilities, drafts exploits, and automatically executes them without human intervention.”
Google said the server held “data exfiltrated from targeted victims” and “multiple scripts capable of facilitating various post-compromise activities, including internal reconnaissance, brute-forcing, installing webshells, and removing other attackers’ webshells.” Google wrote: “Victim telemetry on the server suggests the operator was specifically interested in targeting French websites associated with the media and entertainment industries, as well as the education sector.”
Google wrote that, by chaining these capabilities, “threat actors can dramatically reduce the window between patch disclosure and weaponized exploitation.” It also said LLMs are “shrinking the patch window for organizations.” Google cited a PaperCut case in which, it said, security researchers and the cybersecurity company Huntress “generated working exploits for the vulnerabilities within a day of disclosure.”
In September 2026, according to Google, GTIG analyzed vulnerability statistics and found that overall disclosures and exploits nearly doubled between January and August 2026. In that same analysis, Google said zero-day exploits rose from a 2025 average of eight a month to 11 a month in 2026, then reached 22 in August 2026.
#Which other agent frameworks did Google see in use?
Google says that, in addition to the pentesting tool ARTEX, threat actors are using multi-agent frameworks such as CAIRN, CyberStrikeAI, and PentAGI. Google wrote: “These frameworks share several core capabilities: discovering and exploiting low-complexity vulnerabilities, using LLMs to generate single-use payloads on-demand, and using real-time execution feedback to dynamically adjust tactics.”
Google says GTIG observed CAIRN “used in intrusion activity against targets including in the aerospace sector in the US.” Between late August and early October 2026, Google says, GTIG tracked a threat actor using a customized attack orchestration platform built from PentAGI “to conduct systematic reconnaissance and vulnerability scanning against a wide range of organizations in Asia.” Google also says GTIG has seen experimentation with CyberStrikeAI and HexStrike AI.
Google wrote: “We have observed several threat actors separately experimenting with the use of these frameworks and harnesses, including financially motivated, as well as suspected espionage threat actors with a PRC nexus.” PRC here means the People's Republic of China. Google also wrote: “Given the low barrier to entry, ease of integrating new exploits, and massive scaling potential, we believe it is highly likely that threat actors will continue to use these frameworks in campaigns targeting organizations globally.”
#What did Google report about Claude Code?
In the North Korea section of the same post, Google wrote that it continues to see threat actors from the Democratic People's Republic of Korea (DPRK) using LLM tools to support DPRK IT worker activity, such as generating resumes and preparing for job interviews at Western organizations. According to Google, “a suspected MIDNIGHT NEPTUNE threat cluster combined LLM-enabled social engineering and abused organizations' AI tools as a supply chain attack vector.”
Google said that, in intrusions targeting cryptocurrency organizations in North America and Europe, “the threat actor exploited LLM software architectures as a lateral execution surface.” For initial access, Google said, the actor created fraudulent recruiter profiles on professional networking platforms whose profile imagery contained SynthID watermarks, which Google said confirmed use of LLMs to synthesize deceptive professional identities. Google said those personas directed targets to fraudulent meeting platforms. After that access, Google said, the actor moved into internal code repositories, “modifying .claude/settings.json configuration files so that when users started up Claude Code, it would automatically trigger the download of the SOMBERMEME backdoor.”
Google also said suspected MIDNIGHT NEPTUNE “further masked its lateral movement by generating GitHub Personal Access Tokens (PATs) named after Microsoft Copilot”.
#What does Google say defenders should know?
Google wrote: “While the mechanics of the intrusions remain the same, such as using compromised credentials and targeting unpatched systems, the speed and scale at which frameworks can exploit these known weaknesses introduces a new defensive challenge.”
It also wrote that the frameworks “currently trade stealth for velocity.” According to Google, “Unrestricted autonomous execution generates forensic traffic, relies on recognizable identifiers, and leaves behind distinct code-level traces; while these artifacts increase the likelihood of detection and mitigation, these tools remain highly capable of achieving operational objectives.” Google also said the lower barrier to entry has fueled a wave of LLM-generated exploits that are claimed to work but do not, creating additional verification work for defenders.
Google wrote: “Google Security Operations (SecOps) has deployed specialized behavioral and volumetric rulepacks to detect use of ARTEX.” It also wrote: “Mandiant Consulting released a hardening guide and architectural defense framework specifically designed to counter the speed and scale advantage of adversaries using LLM-enabled offensive tooling.”
#How does this compare?
Google's November 16, 2026 tracker is a separate account from earlier ARTEX coverage on DotsBot, including Amazon's report that ARTEX agents recorded LLM API key theft and Microsoft's report that an AI agent reached Key Vault data in a five-minute case. The Claude Code vector Google describes is also a separate account from our earlier reporting on a malicious Claude Code extension on the Cursor marketplace and on Claude Code approval files that exposed secrets in public GitHub copies.
Our analysis: Google's account is of a pipeline that, Google says, compared patches, drafted exploits, and ran them without a human in the loop, plus other frameworks Google says it has seen in live intrusions, and a suspected cluster that changed Claude Code settings in internal repositories. Google's own defensive points are that patch windows are shrinking, that the frameworks are currently noisy enough to detect, and that Mandiant has published a hardening guide. Organizations that use Claude Code should treat unexpected changes to .claude/settings.json in repositories like code changes and review them before they are merged. That review advice is ours, not Google's.
Frequently asked questions
What did Google say exploit_agent did?
Google says GTIG found an exposed staging server hosting exploit_agent, a multi-agent framework that integrates with multiple frontier and local LLMs. Google wrote that it compares code before and after patch announcements, analyzes findings, drafts exploits, and automatically executes them without human intervention.
Did Google say this was the first time attackers used AI this way?
Yes, as a claim. In its November 16, 2026 tracker, Google wrote that GTIG identified threat actors using LLM frameworks “to fully automate vulnerability discovery and exploitation without a human in the loop for the first time.”
Which other frameworks did Google mention, and who was targeted?
Google says threat actors are also using CAIRN, CyberStrikeAI, and PentAGI, in addition to ARTEX, and that it has seen experimentation with HexStrike AI. Google says GTIG observed CAIRN used against targets including in the US aerospace sector, and tracked a customized PentAGI platform used against a wide range of organizations in Asia.
What did Google report about Claude Code?
In the North Korea section of the post, Google said a suspected MIDNIGHT NEPTUNE threat cluster modified .claude/settings.json files in internal repositories so that starting Claude Code downloaded the SOMBERMEME backdoor. Google placed this in intrusions targeting cryptocurrency organizations in North America and Europe. The North Korea link is Google's assessment, not an independently established fact in this article.
What defensive measures does Google describe?
Google says these frameworks shrink the window between patch disclosure and weaponized exploitation, and that they currently trade stealth for velocity, which can make them detectable. Google Security Operations has deployed rulepacks to detect ARTEX, and Mandiant Consulting released a hardening guide for LLM-enabled offensive tooling.
Sources
Every factual claim in this article traces back to one of these.
One email a week on AI agents and bots.
Launches, pricing and rollout changes and the guides worth reading. At most once a week, and you can leave with one click.
