Google’s Threat Intelligence Group says it has, for the first time, found attackers using large language model frameworks to fully automate vulnerability discovery and exploitation, without a human in the loop.
In a 16 November report, GTIG wrote: “In the past quarter, GTIG identified threat actors using LLM frameworks to fully automate vulnerability discovery and exploitation without a human in the loop for the first time.” A same-day Keyword post, dated 16 November 2026 at 18:30 GMT and headed “The era of autonomous attacks has arrived,” points readers to that report. GTIG has not named the operator behind the framework it describes.
What GTIG says exploit_agent did
“GTIG identified an exposed staging server hosting a multi-agent vulnerability research and weaponization framework, dubbed exploit_agent, that integrates with multiple frontier and local LLMs.” The framework, GTIG said, “compares code before and after patch announcements, analyzes findings to identify potential vulnerabilities, drafts exploits, and automatically executes them without human intervention.”
Victim telemetry on that server, the report says, suggests the operator was specifically interested in targeting French websites associated with the media and entertainment industries, as well as the education sector. Configuration comments on the server were in Russian and included the phrase “exploit_agent pilot,” according to GTIG. That is a description of the comments the researchers showed, not an identification of who ran the system.
GTIG also set the finding against a wider shift in the numbers. It said overall disclosures and exploits “nearly doubled between January and August 2026”, while zero-day exploits rose “from a 2025 average of 8 per month to 11 per month in 2026—before spiking to 22 in the month of August.” The same report names other multi-agent offensive frameworks — ARTEX, CAIRN, CyberStrikeAI and PentAGI — and activity it groups under LLMJacking and “LLM Mixers.” Those are separate strands in the same tracker, not additional names for exploit_agent.
The report also cites what it calls open-source corroboration of fully automated exploitation pipelines. According to GTIG, GreyNoise reported a campaign after two PaperCut print-management zero-days were disclosed in late August 2026, and PHP vulnerabilities that were discovered using LLM-enabled analysis had official patches on 21 October 2026. Those details belong to GTIG’s account and to the firms it cites; they are not, in the report, presented as the exploit_agent staging server itself.
Google’s Keyword post adds a short product note: the company says it is offering defenders Google AI Threat Defense, which it describes as combining Gemini and other frontier models with Wiz, CodeMender and Mandiant. That is a product mention, not a claim that this architecture has stopped the activity GTIG describes.
The finding sits in a line of TSN files on AI tools being used in, and against, cyber operations. The FBI has issued a FLASH on the ARTEX framework, and Microsoft has described AI agents trying 2,250 tenants after a ConnectWise breach. On the defensive side, OpenAI has launched GPT-6 Cyber for defensive research, and Google has given Gemini 4 Carbon to cyber defenders before developers. Microsoft has also said North Korean hackers use AI to scale familiar attacks. Those are different cases. GTIG’s new claim is narrower: a first instance, in the past quarter, of fully automated discovery and exploitation with no human in the loop.
A first automated pipeline is not a named operator
GTIG’s most striking sentence is a claim about what its researchers identified in the past quarter, not a census of every attacker. The evidence it puts forward is an exposed staging server running a framework it has dubbed exploit_agent. That is a significant threshold if the description holds, and it is still one server, described by the company that found it.
The Russian-language comments should be kept in that box. GTIG showed comments that include “exploit_agent pilot.” It has not named an operator, and the comments leave that gap unfilled. A configuration note in a given language is not an attribution to a state, a group or a region.
The supporting figures and the cited campaigns are also easy to flatten, and they should not be. “Nearly doubled” refers to overall disclosures and exploits between January and August 2026. Zero-day exploits, on GTIG’s own split, rose only from a 2025 monthly average of eight to 11 in 2026, before a spike to 22 in August. GreyNoise’s PaperCut campaign and the PHP vulnerabilities are presented as corroboration of a wider trend, attributed to those firms as GTIG cites them, not as a second look at the same staging server. Until GTIG or another official channel names who ran exploit_agent, or shows how widely it was used beyond the French-site telemetry it records, the report is a first documented case of a fully automated pipeline — and not yet a map of who is using one.
Sources
- Google Cloud Threat Intelligence, “Agentic threat activity crosses a significant threshold,” 16 November 2026
- Google Keyword, “The era of autonomous attacks has arrived,” 16 November 2026
Get TSN’s biggest tech stories, explained.
News and insights across AI, tech, crypto and cybersecurity. Every source linked. Unsubscribe anytime.
Get the newsletterRelated stories
- FBI FLASH Sets Out ARTEX Mitigations and Indicators is the closest recent TSN file on a named multi-agent offensive framework, and the right next stop for that official advisory.
- OpenAI Launches GPT-6 Cyber for Defensive Research
- Microsoft Says AI Agents Tried 2,250 Tenants After ConnectWise Breach
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
