China-nexus threat actors unleashed another wave of attacks using autonomous AI tools to compromise microprocessor manufacturers, US defense contractors, and other organizations.

The malicious activity, tracked as CL-STA-1238 by researchers at Palo Alto Networks’ Unit 42, is the second documented attack in which China-linked actors used the open source AI penetration-testing framework called Cairn for automation, they revealed in a blog post. In this latest campaign, attackers conducted 25 operations between August and September against at least 19 organizations, achieving “unauthorized access or data retrieval at 10 of the targeted organizations,” according to the post.

In the successful attacks, the threat actors obtained dozens of valid credentials; achieved mass exposure of personal data, exceeding 1 million records at one organization; retrieved military and government identities; created persistent administrator accounts in a production system; and modified a live industry standards-ballot document.

Related:Inside an AI Attack Against a US Retailer

Of the total organizations targeted, seven were in the semiconductor and advanced-manufacturing supply chain, and three organizations were in US defense and defense policy. The attackers also targeted various human rights and civil society organizations.

According to Unit 42 researchers, the victims included a industry standards body, which suffered a password-spraying attack; a semiconductor company, where attackers brute-forced a production credential on a partner-integration application; a US teacher’s union, which suffered an unauthenticated account-creation bypass on a content management system; and a defense policy think tank, where attackers targeted an open back-end registration endpoint on its identity platform.

The China-nexus actors also targeted content management systems through identity and configuration weaknesses at other organizations, including a semiconductor-lithography manufacturer, an industrial-automation manufacturer, and an education nonprofit.

Shift in AI Offensive Maneuvers

Cairn played a significant hand in the campaign, and the use of it “represents a shift in how attackers are applying AI to targeted offensive operations,” according to the post.

“The attackers supplied a target and a goal, then Cairn planned, executed, and recorded results on its own,” Unit 42 researchers wrote.

Cairn is AI model agnostic, and in this case, attackers weaponized it using Chinese large language models (LLMs) Zhipu GLM 5.2 and DeepSeek V4 Flash and Pro. However, initial exposure to the organizations’ networks came through many of the traditional security gaps found across industries, the researchers noted.

Related:Artex: The AI Pen-Test Tool Hackers Turned Into a Weapon

“It doesn’t matter much to a victim whether a human or an AI agent is behind the keyboard,” Tyler Oliver, CEO and founder of threat-hunting platform Huntbase, tells Dark Reading. “A human might hand the tool a mission and walk away, or direct it step by step. Either way, the password spraying looks like password spraying.”

Indeed, the vulnerabilities the attackers exploited were known, according to Unit 42, including “authentication bypasses; insecure direct object references; mass-assignment privilege escalation; weak and default credentials; exposed cloud APIs; and dated public CVEs.”

AI Agents Target Set Objectives

As in the previous Cairn-driven attacks, the findings once again provide a look under the hood about how threat actors can weaponize the tool against organizations, a blueprint that may be used in future attacks.

Rather than manually directing every stage of an attack, operators supplied Cairn with a target and a high-level objective. The framework then developed its own plan, conducted reconnaissance, identified potential vulnerabilities, attempted exploitation and recorded its findings.

Related:CISA Announces Enhanced Reporting Form for Cyber Incidents

“An AI model drives each Cairn ‘agent worker,'” according to the post. “Cairn coordinates its activities through a shared ‘blackboard.’ Agents read the blackboard and write their findings back as a fact-intent graph, while human operators can inject ‘hints’ at any time to help guide the operations if needed.”

“Cairn has a shared record of what it’s already learned, so it can build on earlier findings without a person reviewing every result and deciding what happens next,” Rickard Carlsson, CEO of AI security firm Detectify, tells Dark Reading. Meanwhile, this ability to work with limited supervision is already showing up across multiple tools including another AI penetration-testing framework, Artex, which also is being weaponized by attacker, he says.

Individual CL-STA-1238 operations lasted anywhere from 20 minutes to more than 50 hours, according to Unit 42. Moreover, attackers could achieve simultaneous operations against multiple organizations, which reduced the dependence on human operators to manage each intrusion.

Old and New Defense Tricks

Though AI is changing the game in terms of speed, volume, and expanse of attacks, it does not change the “complexity” of attacks, which, as mentioned, often rely on known security gaps, according to Unit 42.

That’s a spot of good news for defenders, because “identity hygiene, attack-surface reduction, and software patching remain effective” against attacks that use Cairn or other AI tools, according to the post.

However, there are other safeguards that organizations can put in place to defend against AI-driven attacks, too, says John Strand, owner of security consulting firm Black Hills Information Security.

“There are two things I think every organization should be doing right now,” he tells Dark Reading. “First, continuous attack surface management. Organizations need to identify all the different services and systems they have exposed to the internet and start restricting access wherever possible.”

The second key defense move is to start using cyber deception as much as possible by setting up “tripwires” that can detect attackers during reconnaissance and lateral movement “before they accomplish their objectives,” Strand adds.





Source link

#

No responses yet

Leave a Reply

Your email address will not be published. Required fields are marked *