Secure Bulletin Navigating the cyber sea with knowledge
Home > Articolo > Google Warns Cybercriminals Have Crossed a ‘Critical Inflection Point’ With Autonomous AI Attack Agents
Google Warns Cybercriminals Have Crossed a ‘Critical Inflection Point’ With Autonomous AI Attack Agents
Read Time:4 Minute, 16 Second

Cybercriminals, state-backed spies, and influence operators have crossed a line that security researchers have been warning about for years: they are no longer just using AI chatbots for occasional assistance, but are running autonomous AI agents that scan networks, break into systems, and move laterally through infrastructure with minimal human supervision.

That’s the central warning in Google Threat Intelligence Group (GTIG) and Mandiant’s latest AI Threat Tracker, published November 16, 2026. The report describes AI misuse in cybercrime, espionage, and influence campaigns as having reached a “critical inflection point,” drawing on real Mandiant incident response cases, ongoing tracking of threat actor groups, and telemetry from Google’s own defensive platforms.

A Toolbox Full of Offensive AI Frameworks

According to the report, threat actors are actively deploying several open-source, multi-agent hacking frameworks in live operations. GTIG confirmed real-world use of two frameworks in particular: one called ARTEX, which translates an operator’s high-level goals directly into executable attack commands and has already been used to automate Active Directory attacks in bursts lasting as little as two to 25 seconds; and another called CAIRN, which coordinates multiple specialized AI agents through a shared workspace using an observe-orient-decide-act loop, and was observed navigating attack paths against targets in the US aerospace sector.

A third framework, PentAGI, which breaks attack objectives into smaller automated tasks, was found running in a customized form against a wide range of organizations across Asia between late August and early October 2026, conducting reconnaissance and vulnerability scanning largely unattended. Google also flagged active testing — though not yet confirmed operational deployment — of two additional platforms: CyberStrikeAI, which links AI models to more than 100 security tools and ready-made attack scripts, and HexStrike AI, which connects to over 150 offensive tools through the Model Context Protocol for automated reconnaissance, web application testing, and exploitation.

Separately, security firm Wiz tracked an ARTEX-wielding actor overlapping with the group Microsoft calls Storm-0940 that targeted identity platforms including Okta, Microsoft Entra, and Amazon Cognito at IT and software vendors. Google notes these tools generally sacrifice stealth for raw speed — leaving behind noisy traffic patterns and recognizable code — but they are still proving effective enough to reach their objectives regardless.

Exploits Written and Launched Without a Human in the Loop

Perhaps the most alarming finding is one Google describes as a first: GTIG identified attackers using AI to discover and weaponize software vulnerabilities with zero human involvement in the exploitation chain itself. Investigators found a framework, nicknamed “exploit_agent,” sitting on an exposed staging server. It worked by automatically comparing a web application’s code before and after a security patch was applied, using the difference to identify the underlying flaw, then writing a working exploit and firing it off entirely on its own. Russian-language comments found in the tool’s configuration files suggest it may still have been in a pilot phase, with its operator apparently focused on French media, entertainment, and education websites.

This isn’t a theoretical risk. Threat intelligence firm GreyNoise previously documented an AI-orchestrated campaign that exploited two zero-day flaws in PaperCut print management software, striking hundreds of organizations and reaching full domain administrator access within minutes of the initial compromise. Google’s own data shows the broader trend accelerating sharply: zero-day exploitation climbed from an average of 8 incidents per month in 2025 to 11 per month through most of 2026, before spiking to 22 in August alone.

Stolen AI Access Becomes Its Own Black Market

Financially motivated groups are also industrializing the theft and resale of AI credentials and cloud compute access through a scheme researchers call “LLM Mixers.” Google’s related LLMjacking research uncovered cases where fraudulent usage charges topped $100,000 in a single day, with some individual victims facing bills exceeding $1 million. Attackers reportedly test stolen API keys with small, inconspicuous requests first, specifically to avoid tripping billing alerts before scaling up abuse.

On the espionage front, the report describes a suspected North Korean group, tracked as MIDNIGHT NEPTUNE, that built fake recruiter personas using AI-generated profile photos to lure cryptocurrency firm employees into “ClickFix” infection chains. Once inside, the group modified local Claude Code configuration files so the tool would automatically fetch a custom backdoor at startup, while disguising its infrastructure access tokens to look like ordinary Microsoft Copilot activity.

What Defenders Should Do

Google’s Mandiant team recommends a layered response: harden administrative systems, move to short-lived credentials wherever possible, enforce the principle of least privilege, segment networks to limit lateral movement, and closely monitor for sudden spikes in API calls or identity-related activity that could signal autonomous reconnaissance. For organizations running their own AI infrastructure, the report urges setting hard budget alerts, continuously scanning source code repositories for exposed API keys, and treating any unexpected activation of an AI service as a serious security event rather than a billing anomaly.

As the report’s authors put it, the underlying attack techniques themselves haven’t fundamentally changed — AI is simply letting adversaries chain familiar tactics together at a speed that leaves defenders far less time to notice and respond.

Share: Twitter  |  Facebook  |  LinkedIn
Join the discussion

This is a blog in the Fediverse: you can find this article everywhere with @blog@securebulletin.com and every comment/answer will appear here.

If you want to comment on Google Warns Cybercriminals Have Crossed a ‘Critical Inflection Point’ With Autonomous AI Attack Agents, use the discussion on Forum.

>> forum community

Comments

Leave a Reply