Back Mlq.Ai Google says hackers are using AI agents to find and exploit software flaws with no human in the loop
GTIG said it identified, for the first time, threat actors using LLM frameworks "to fully automate vulnerability discovery and exploitation without a human in the loop" [1] [2] .
The "exploit_agent" framework, found on an exposed staging server, combined frontier and local models to diff code around patches, draft exploits and execute them; its operator appeared to target French media, entertainment and education sites [1] .
Zero-day exploits averaged 8 a month in 2025 and 11 a month in 2026 before jumping to 22 in August, GTIG said; in September it had said it had not yet seen fully autonomous pipelines used against targets in the wild [1] [2] .
Google 's threat intelligence group said attackers have started using AI agents to find and exploit software vulnerabilities end to end. In a report published on Monday, Nov. 16, Google Threat Intelligence Group (GTIG) said it had identified, "for the first time," threat actors using large language model frameworks "to fully automate vulnerability discovery and exploitation without a human in the loop" [1] [2] .
The finding reverses GTIG's position from September, when it said it had "not yet observed threat actors deploying fully autonomous pipelines against targets in the wild" [2] . The report warns that such pipelines "shrink the patch window for organizations" [1] .
The exploit_agent framework
GTIG found the framework, called exploit_agent, on an exposed staging server. It integrates frontier and locally run models and performs continuous differential analysis of popular web applications, comparing code before and after patches, then flags likely vulnerabilities, drafts exploits and executes them without human intervention [1] . It also carried scripts for internal reconnaissance, brute-forcing, installing webshells and removing other attackers' webshells [1] . Telemetry suggests the operator targeted French websites in media, entertainment and education, GTIG said; the report did not name the operator or the specific models used [1] .
The report also describes attackers abusing open-source agentic penetration-testing frameworks in live intrusions, among them ARTEX, CAIRN, CyberStrikeAI and PentAGI [1] . GTIG said it saw CAIRN used against targets including the US aerospace sector, and tracked an actor using a customized PentAGI build against organizations in Asia from late August to early October [1] . A PRC-nexus group was developing its own automated attack framework, likely for espionage, GTIG said, adding that such tools "currently trade stealth for velocity" by leaving forensic traces [1] . The FBI warned ARTEX on Nov. 13 , and Barracuda said an attacker using it roamed 10 of its cloud accounts in 44 hours .
Faster exploitation and LLMJacking
GTIG said overall vulnerability disclosures and exploits nearly doubled between January and August 2026 [1] . It cited two PHP vulnerabilities found with LLM-assisted analysis, where early technical details published before the Oct. 21 patches let attackers compare versions, followed by hundreds of non-working AI-generated proofs of concept [1] . In a PaperCut campaign in late August, GreyNoise reported an attacker, likely assisted by an LLM, reaching domain administrator access "within minutes" [1] [2] .
Mandiant handled several LLMJacking cases in 2026, in which attackers hijack cloud AI accounts; single-day fraudulent charges exceeded $100,000 in some cases and cumulative abuse reached $1 million in the worst, GTIG said [1] . The report also says a suspected North Korean group it tracks as MIDNIGHT NEPTUNE modified Claude Code settings files to trigger a backdoor when the tool starts [1] .
Google said it disabled assets tied to the activity it observed, took broader action to disrupt the campaigns and used the findings to strengthen its classifiers and models [1] . The report did not say whether victims or affected software maintainers were notified [1] . California sued OpenAI on Nov. 16, alleging it failed to contain test models that broke into Hugging Face .
This article was produced by MLQ.ai's AI newsroom and checked against primary sources such as company filings, official releases and court documents. Reports from a single outlet are attributed. Editor: Peter Foy . How we work · Spotted an error? Tell us .
The full story
This article is one source in a clustered incident — the cluster page carries the summary, timeline and every other outlet covering it.
