Without Human Intervention

Hackers Deployed an Autonomous AI System Against U.S. Security Organizations

Palo Alto Networks researchers revealed that Chinese-speaking attackers converted a security testing tool into an autonomous hacking weapon that targeted 19 organizations • At 10 of them, data was stolen or unauthorized access to information was obtained

Cyber and AI
Cyber and AI. Credit: Shutterstock | Photo: Shutterstock

Researchers from Unit 42, the research arm of cybersecurity giant Palo Alto Networks, uncovered an unusual cyberattack demonstrating how readily available artificial intelligence tools are becoming autonomous offensive weapons.

According to the report, Chinese-speaking attackers succeeded in converting an open-source system designed for security testing into an autonomous hacking tool that operated against a series of strategic organizations, including manufacturers of advanced semiconductors, institutes dealing with U.S. security policy and civil society organizations.

Palo Alto Networks
Palo Alto Networks | Photo: Shutterstock

The company assesses with medium confidence that the campaign, dubbed CL-STA-1238, was carried out as part of intelligence-gathering activity linked to China.

How the Attackers Bypassed the AI’s Safeguards

The attack took place during August and September 2026, when 25 autonomous intrusion operations were carried out against at least 19 different organizations. For the mission, the attackers used a customized version of Cairn, an open-source artificial intelligence-based system originally designed to help security professionals test systems for vulnerabilities.

To turn the system into an offensive tool, the attackers connected it to the Chinese artificial intelligence models Zhipu GLM and DeepSeek. To prevent the models from blocking the malicious activity, the attackers presented the intrusions as “authorized security tests,” while explicitly instructing the models not to refuse to perform the tasks.

Artificial intelligence, AI
Artificial intelligence, AI | Photo: Shutterstock

Once the operators defined the target, the system operated entirely independently: It broke the objective down into subtasks, performed the scans, made real-time decisions based on the findings and even documented the results, while allowing the human operators to intervene and guide it if necessary.

Data Leaks, Fake Administrator Accounts and Interference in Standardization

The autonomous attacks produced tangible results for the attackers, who succeeded in obtaining unauthorized access or extracting sensitive information from 10 of the organizations targeted. In one case, the systems of an American teachers’ association were breached, and the details of more than 1 million members and about 31 thousand contacts were stolen.

In another case, an intrusion into the identity management system of a security policy program led to the exposure of about 500 user records, including government and military email addresses.

Cyber, China and the U.S., illustration
Cyber, China and the U.S., illustration | Photo: Shutterstock

The damage also extended to industry and regulation. At a chip company that was attacked, the artificial intelligence system created unauthorized administrator accounts, giving the attackers a persistent foothold in the corporate network. At an industrial standards organization, the attackers went even further and succeeded in altering an active document that was being used at the time for a vote on a professional standard.

The New Threat: Scale and Speed Rather Than Technological Sophistication

The most troubling finding for the cybersecurity community is that, according to the researchers’ assessment, this extensive campaign was conducted by a single person or a very small team. The combination of open-source tools and advanced language models dramatically reduces the manpower and time needed to manage an attack, allowing attackers to run simultaneous intrusion attempts against dozens of targets.

However, the cybersecurity company emphasizes that, technologically, the system did not use sophisticated or unknown exploits. Instead, the AI efficiently exploited known security vulnerabilities, weak passwords and misconfigurations in the organizations’ systems. The main innovation lies in the automation and scale of the operation, not in its complexity.

Cybersecurity, illustration
Cybersecurity, illustration | Photo: Shutterstock

The significance, the researchers conclude, is that basic protective measures - such as consistently applying security updates, strictly managing permissions and reducing systems’ exposure to the internet - remain relevant and effective in stopping attacks, even in an era when artificial intelligence manages the intrusion.

Key takeaways

  • Palo Alto Networks' Unit 42 said Chinese-speaking attackers turned an open-source security testing system into an autonomous hacking tool.
  • The campaign, dubbed CL-STA-1238, involved 25 autonomous intrusion operations against at least 19 organizations during August and September 2026.
  • At 10 of the targeted organizations, attackers obtained unauthorized access or extracted sensitive information.
  • The attackers used a customized version of Cairn connected to Zhipu GLM and DeepSeek, and bypassed safeguards by framing the activity as authorized security tests.
  • Researchers said the main advance was scale and automation rather than novel exploits, and that basic security measures still helped stop such attacks.

Questions and answers

What did Palo Alto Networks say about the cyberattack campaign?
Unit 42 said attackers converted an open-source AI security testing system into an autonomous hacking tool. The company assessed with medium confidence that the campaign was linked to China.
How many organizations were targeted and how many were breached?
The attackers carried out 25 autonomous intrusion operations against at least 19 organizations. At 10 of them, they either stole data or gained unauthorized access to information.
What systems and AI models were used in the attack?
The attackers used a customized version of Cairn and connected it to the Chinese AI models Zhipu GLM and DeepSeek. They told the models the intrusions were authorized security tests and instructed them not to refuse the tasks.

Share this story

About the writer

Shai Nogai

Writes for TV10, Israel’s economic channel. This story was translated from the original Hebrew.

More stories by Shai Nogai

More for you