Attackers Use Autonomous AI And Target Semiconductor, Defense, Civil Society Sectors

Unit 42 researchers revealed on November 12, 2026, that Chinese-speaking attackers used an open-source AI penetration testing framework called Cairn to autonomously conduct 25 operations against at least 19 organizations. The attackers weaponized Cairn by pairing it with Chinese AI models Zhipu GLM and DeepSeek. The attackers supplied a target and a goal, then Cairn planned, executed and recorded results on its own.

These attacks achieved unauthorized access or data retrieval at 10 of the targeted organizations. They accessed dozens of valid credentials, exposed massive amounts of personal data (exceeding 1 Mil records at one organization), retrieved military and government identities, created persistent admin accounts in a production system, and modified a live industry standards-ballot document.

Darin Fredde, Senior Director and Offensive Security Evangelist, Ridge Security

“In attacker does not need your AI to misbehave. They can bring their own. Cairn is one example in a growing ecosystem of publicly available AI security tools. The same capabilities that help an organization uncover an exposed application can help an intruder pursue it. My concern is that treating the availability of these tools as the problem could leave legitimate defenders with fewer ways to test themselves, while determined attackers modify what is already available.

“We should preserve proactive defense. Organizations need opportunities to discover how their systems fail before someone else does. But access to a powerful testing tool also creates responsibilities: define its scope, constrain its privileges, and protect the information it collects. A penetration test can accumulate credentials, sensitive records, and a detailed map of an organization’s weaknesses. Its dashboard and evidence store become sensitive assets themselves. Attackers have no obligation to preserve those protections.

“In the campaign Unit 42 describes, operators used a modified framework and fabricated authorization claims that the models accepted. An authorization file written by the attacker is hardly permission from the victim. That is why responsible-use language and model guardrails cannot carry the entire security burden. They can help, but organizations need controls that continue to enforce boundaries when someone ignores the instructions, changes the model, or modifies the tool. The urgency comes from what happens after discovery. A weakness that once required an operator’s sustained attention can now become part of an automated search for access. Defenders should use that same opportunity to test exposed systems, fix exploitable paths, and verify that the fixes hold.

“I would resist broad restrictions that weaken legitimate testing without a demonstrated reduction in abuse. Give defenders the ability to investigate aggressively and require evidence that their tools respect scope and protect what they discover. Meanwhile, build enterprise defenses on the assumption that an attacker’s agent will keep trying, regardless of what its acceptable-use policy says.”

Waseem Ahmed, Head of Engineering, Secure.com

“This is a real shift. With tools like Cairn, one operator gives AI agents a one-line goal, and they handle recon, scanning, and exploit attempts. In one campaign Reuters reported, a researcher estimated the AI cost at roughly $50 per successful hack. Unit 42 says autonomous AI raises the volume and coverage of attacks, not their complexity, so the weaknesses are old ones.

The agents use normal tools, so watch behaviour, not files: bursts of scans and logins at machine speed, and access that hops between systems. Unit 42 links a separate Cairn campaign, with moderate confidence, to China-nexus intelligence collection. It targeted at least 19 organisations, including chip makers, defense and policy groups and civil society, and breached 10.

“When we aimed an AI pentest pipeline at three live systems over one weekend, it found 21 flaws, seven of them critical. That was on systems we own. Test yourself with AI first, fix fast, then test again.”

The fact that this was written up means that this is a “today” problem and should be treated as such.

Leave a Reply