Hackers used a fast-emerging, AI-backed hacking framework in an intrusion targeting the US aerospace sector, Google's threat intelligence arm said Monday in findings reported by Nextgov. The tool, known as Cairn, is a publicly available penetration-testing framework tied to the Chinese open-source security community — and its appearance inside a live intrusion marks another step in the migration of agentic AI from research demos into real espionage operations.

Google did not name the aerospace organization that was accessed, say what information may have been taken, or attribute the intrusion to a particular government. The report broadly covers activity observed over the last quarter, and it lands amid today's AI industry coverage of agentic systems being repurposed faster than defenders can adapt.

What Cairn and Its Cousins Can Do

Cairn differs from the chatbot-assisted hacking that dominated earlier threat reports. Rather than a human prompting an AI assistant step by step, Cairn uses AI agents that share information with one another and decide which hacking techniques to try next as they work toward penetrating a system.

Google researchers said other open-source systems — including ARTEX, CyberStrikeAI and PentAGI — are now being used to "execute multi-stage intrusions at greater operational velocity" compared with earlier activity in which hackers manually prompted AI chatbots to assist with individual steps. ARTEX was separately flagged in an FBI cybersecurity advisory posted Friday.

The Most Alarming Finding: Fully Automated Exploitation

One discovery stands out from the quarterly report. Google researchers found an attacker's system configured to study software fixes, develop attacks based on what it learned, and execute those attacks automatically — with no human involved in the loop. The company said it was the first time its researchers had identified attackers fully automating that process.

The implication is stark: the race between patch releases and exploitation could collapse from weeks into hours, putting further pressure on organizations that already struggle to install security updates before criminals reach them. The Hacker News, covering the same findings, summarized the technique as hackers using AI agents to build and run exploits directly from software patches.

"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," the company said, according to Nextgov.

Chinese Teams Industrialize Analysis; North Korean Recruiters Go Synthetic

The report also details how state-linked actors are folding models into different stages of their operations. Chinese hacking collectives have used AI models to analyze stolen files and produce finished intelligence reports for customers. In one instance, a contractor used AI to translate and analyze stolen information, work out how targeted organizations were structured, and organize the findings into intelligence reports — using the technology for the complex interpretation work that follows a theft rather than the theft itself.

Google also observed suspected Chinese operatives experimenting with automated hacking tools, including one China-linked group developing its own system, likely for use in spying operations.

North Korean actors, meanwhile, are using AI to sharpen their social engineering. One group Google tracks targeted cryptocurrency organizations in North America and Europe using fabricated recruiter profiles. The fake recruiters directed targets to counterfeit meeting services, where instructions to fix a supposed technical problem led them to run malicious code.

Once inside, the attackers changed settings for Claude Code — Anthropic's AI coding assistant — in company code repositories, so that launching the assistant would trigger a malware download. Weaponizing a trusted developer tool means the malicious action rides on software engineers already trust.

The findings echo a Justice Department announcement last week targeting FaceStealer, a North Korean fake-video-call scheme linked to more than $400 million in theft, alongside court orders covering more than $1.2 billion stolen from Binance. Google's report did not connect its findings to either case.

Influence Operations Get More Automated Too

Beyond espionage and financial crime, Google said influence operations linked to Russia, China and Iran now routinely use AI models to produce content and automate its distribution — delegating more of the work of posing as ordinary people online to machines.

A Widening Gap

Taken together, the findings sketch a defensive problem that grows with every open-source release. Penetration-testing frameworks like Cairn, ARTEX, CyberStrikeAI and PentAGI are legal tools built by and for the security community, and their agentic architecture makes them genuinely useful for red teams. The same properties make them turnkey offensive weapons for anyone who downloads them.

Google's message to defenders is effectively a planning assumption rather than a prediction: assume the next intrusion you face was executed at machine speed, on a timeline set by an agent that read your patch notes before you finished deploying them.

---

Stay Ahead of AI

Get the latest AI news, analysis, and breakthroughs — all in one place.

Read more AI news →