Google says attackers now run autonomous multi-agent frameworks
GTIG documented a framework, exploit_agent, that diffed patches, wrote exploits and ran them with no human, which it called a first; zero-day exploits averaged 11 a month in 2026.
- Security & misuse
- Notable
Google’s Threat Intelligence Group (GTIG) and Mandiant published their quarterly AI threat report on 16 November, titled “Agentic Threat Activity Crosses a Significant Threshold”. It said adversaries have moved beyond prompting models to deploying autonomous multi-agent frameworks and automated exploit pipelines, and that the “adversarial AI landscape has reached a critical inflection point”.
On frameworks, GTIG named ARTEX, CAIRN, CyberStrikeAI and PentAGI as open-source tools used by threat actors, and said it had independently seen CAIRN and PentAGI in the wild, CAIRN against targets including US aerospace. It said the tools generate single-use payloads on demand and adjust to execution feedback, but “trade stealth for velocity”: they leave forensic traffic and code-level traces that raise the chance of detection. The FBI had issued its ARTEX advisory three days earlier, and Amazon reported on Cairn four days later.
The report’s most specific claim concerned vulnerability research. GTIG said it identified, “for the first time”, threat actors using LLM frameworks to automate discovery and exploitation with no human in the loop. It found an exposed staging server running a framework dubbed exploit_agent, which compares code before and after patch announcements, drafts exploits and executes them automatically, integrating with several frontier and local models. Victim data on the server suggested the operator was targeting French media, entertainment and education websites. An environment file on the server carried Russian-language comments. GTIG said attackers may prefer weaponising patched n-day bugs, which give a finite set of material to analyse, over hunting zero-days. Zero-day exploits rose from an average of 8 a month in 2025 to 11 in 2026, with 22 in August, while total disclosures and exploits nearly doubled between January and August. GTIG also noted a flood of LLM-generated exploits “claimed to work but don’t”, adding verification work for defenders.
Other findings covered resale of stolen model credentials through proxy services such as “LLM Mixers”, the LLMjacking Mandiant had described earlier in the month, routine use of LLMs by state-linked and commercial influence operations, and manipulation of coding-assistant configurations as a supply-chain vector. GTIG cautioned that better content from such operations does not necessarily mean greater impact. The report is a vendor’s telemetry and assessments rather than independent measurement.
Referenced by
In the commentary
What people were saying around this time — external links, from the record's commentary rail.