# Palo Alto reveals: AI agents accelerate cyberattacks

- Published: 2026-09-22T13:37:35.000Z (Tuesday, September 22, 2026)
- Updated: 2026-11-16T11:13:52.788Z
- By: Alma Tal Tzur
- Section: AI
- Canonical: https://tv10.news/news/tech/ai/115528/palo-alto-ai-agents-accelerate-cyberattacks/
- Hebrew original (TV10): https://tv10.co.il/news/tech/ai/115528/
- Publisher: TV10 Global (TV10, Israel’s economic channel). Translated from Hebrew with AI and an automated accuracy check.

[Palo Alto](https://tv10.news/news/tech/112708/palo-alto-acquires-ai-startup-for-half-billion-dollars/) Networks reported today (Tuesday) that researchers at Unit 42, the company’s research unit, identified how using several AI agents simultaneously allows attackers to dramatically shorten the time needed to breach organizations. In the case examined, the agents carried out more than 50 attack actions in less than 10 hours — a scale that researchers said could have required about two weeks for a coordinated human team.

During the intrusion, the agents moved between different layers of the target environment, including mapping the network and internal systems and locating access credentials in code repositories. This enabled them to obtain administrator privileges, as well as keys used by the organization to operate cloud infrastructure and AI systems.

![](https://tv10.news/media/2026/11/a6/a6b776dfef4c65385c0ec4a3be189f3a33578db0d85dbb3c4b50cfdaa3863883.jpg)

Sam Rubin, senior vice president at Palo Alto Networks’ Unit 42 research unit | Photo: Palo Alto Networks

The intrusion also allowed the attacker to turn the organization’s own infrastructure into part of the attack apparatus. The agents used the victim’s computing resources and AI systems, and one of them even produced a detailed 80-page technical report at the end of the process, documenting the dozens of vulnerabilities that were exploited.

According to the researchers, the attack’s power did not stem from the discovery of a new zero-day vulnerability or the use of a previously unknown technique. The main change was in how the work was carried out: Several agents operated side by side, divided tasks among themselves, reviewed the findings and adjusted the next steps according to the intrusion’s progress.

![Palo Alto Networks](https://tv10.news/media/2026/11/37/3790745f15c97af33d1028baeafc6ee1114636b5c98878ebf374b6ac0d8ff183.jpg)

Palo Alto Networks | Photo: Shutterstock

### On the defensive side as well

The research comes alongside the launch of a new Unit 42 service that seeks to use the same technology for defense. The service deploys AI agents in a controlled environment, continuously simulating intrusion attempts and searching for ways a real attacker could breach an organization.

The system incorporates GPT-5.6-Cyber, Claude Mythos 5 and additional models. The tests are conducted continuously, so when an organization changes its technological environment, the service can resimulate attack scenarios and examine whether new pathways have been created through which its systems can be breached.

![Anthropic’s Claude](https://tv10.news/media/2026/11/a8/a8c55db344b2dbbe38e99bf23e0a4ab81ecb6651bf218ca2fa08f64cd0bdb6b5.jpg)

Anthropic’s Claude | Photo: Shutterstock

The goal is not merely to produce a list of security flaws, but to examine how they can be combined. The system attempts to build a complete attack path that can begin with a vulnerability in an application and advance through cloud services, code repositories, identities and the corporate network.

### Each model sees a different picture

Palo Alto Networks’ tests also point to a limitation of deploying a single AI model. When the models were tested separately, none managed to identify more than 40% of the vulnerabilities in the environments examined. At the same time, the overlap between the findings of GPT-5.6-Cyber and Claude Mythos 5 was below 10% — meaning the two models tended to uncover different vulnerabilities.

![ChatGPT](https://tv10.news/media/2026/11/ba/ba4db81eff90d3402df8b50ed8c7188394c15bd764702d39a6fdac4a1cca4a62.png)

ChatGPT | Photo: hadigraphics, Freepik

According to the company, the system was tested over six months in more than 100 projects carried out for clients. Of the exposures identified, 37% were classified as severe or critical. In addition, more than two-thirds of the vulnerabilities discovered in third-party applications had no known CVE number — meaning they were not documented as known security vulnerabilities in official databases.

Sam Rubin of Unit 42, a senior vice president at Palo Alto Networks, said the shift to using AI agents is changing the balance of power between attackers and defenders. According to him, attackers’ ability to perform many actions rapidly and simultaneously also requires defense systems to adopt processes capable of operating at a similar pace. Rubin said: "Artificial intelligence has given attackers an asymmetrical advantage over organizations that are still trying to defend themselves at human speed. Modern cybersecurity requires defense at machine speed."
