Zero Networks Joins Anthropic Cyber Verification Program to Test AI Powered Attacks Against Microsegmented Networks

In the Anthropic Cyber Verification Program test, Claude broke into a machine that was deliberately left unprotected by network segmentation. But it could not move to other systems protected by segmentation.

CHARLOTTE, N.C., October 21, 2026--(BUSINESS WIRE)--Zero Networks, a leading provider of Zero Trust security solutions, announced today the results of its participation in the Anthropic Cyber Verification Program. The results showed that an advanced AI model could not move beyond a deliberately exposed machine after gaining administrator access and stealing login credentials. In the test, Zero Networks' microsegmentation stopped the model from breaking into protected systems and accessing protected identities. This kept the attack limited to the machine where it began.

This result addresses an important security question as advanced AI models become better at carrying out attacks on their own: What happens when an AI attacker gets inside a network? Recent research from OpenAI showed that AI agents can find security weaknesses, gain higher levels of access and help compromise Hugging Face's infrastructure. OpenAI researchers have warned that attackers could eventually use similar self-directed AI agents against businesses.

Anthropic accepted Zero Networks into its Cyber Verification Program so the company could carry out more detailed security tests using Anthropic's AI research capabilities, which can be used for both helpful and harmful purposes. The test placed an advanced AI model in a realistic network containing different types of systems. It examined whether the model could use stolen login details and attack methods to move from one system to another.

The model successfully broke into the one machine deliberately left outside Zero Networks' segmentation protections. It gained administrator access and extracted login credentials. However, when it tried to use this access to reach protected systems, segmentation stopped it. During the test, the model tried 18 different attack paths, used 23 offensive security tools and applied 22 techniques from the MITRE ATT&CK framework. It never moved beyond the machine it first compromised.

In some test runs, the model recognized that Zero Networks' segmentation prevented it from reaching other systems and stopped trying additional attack paths.

"AI is going to make attackers faster, more persistent and increasingly autonomous. Security teams aren't going to win by trying to out-click or out-think machines operating at machine speed," said Benny Lakunishok, CEO and co-founder of Zero Networks. "The answer is to build environments where compromising one machine doesn't give the attacker somewhere else to go. That's what this test demonstrated: the attacker got in, had credentials and had every incentive to keep moving. The architecture stopped it."

AI Makes the Attacker Faster. Segmentation Makes Faster Irrelevant.

Advanced AI models are becoming better at finding security weaknesses, stealing login credentials, planning attacks with several steps and working with less human help. This means that checking whether an attacker can break into a single device is no longer enough. The more important question is what the attacker can reach after getting inside.

Zero Networks carried out its test with Anthropic to examine the defensive side of this problem: If a highly capable AI attacker gets into a network and obtains login credentials, can the network's design stop it from moving to other systems?

The test showed that the answer was yes in the environment being tested.

"What OpenAI disclosed should worry people well beyond the security team: autonomous agents finding each other, colluding, and escalating access without a human in the loop," said Dr. Chase Cunningham, a cybersecurity expert who specializes in zero trust architecture. "The infrastructure the world actually runs on was never built to assume the attacker on the other end doesn't get tired. Segmentation is one of the few controls that changes that math regardless of how the attacker thinks or how fast it moves, and this test is a concrete demonstration of why that matters."

Methodology

Zero Networks built a large test lab designed to look like a real business network. It included Windows, Linux and macOS systems, Kubernetes and different types of workloads. The company then used Zero Networks microsegmentation to protect the systems and identities in the environment.

One machine was deliberately left outside the segmentation protections. It served as the control machine and the starting point for the attack.

An advanced AI model was set up as an independent attacker. It had administrator access and no restrictions on how it could try to attack the environment. The test examined whether the model could break into systems, obtain login credentials and move between systems using offensive security tools and techniques.

Claude tried 18 different attack paths using 23 offensive security tools and 22 techniques from MITRE ATT&CK. These included Kerberoasting, extracting credentials from Active Directory and attempting a Pass-the-Hash attack against the main target. MITRE ATT&CK is a standard framework used across the cybersecurity industry to organize the methods and techniques used by real attackers. Trying 22 different techniques showed that Claude was not simply repeating the same attack. It tried a wide range of recognized methods, much like a human red team or a sophisticated attacker might.

Claude successfully broke into ESTI, gained administrator access, extracted login credentials and placed a proof file on the system. However, it could not use this access to move into systems protected by Zero Networks segmentation.

About Zero Networks

Zero Networks makes microsegmentation easy to set up, automate and use. The company helps organizations stop attackers from moving between systems and keep their businesses running even when an attacker gets inside. It does this by automatically creating and enforcing identity-based least-privilege rules for users, workloads and AI agents. These rules use security controls already built into the environment, without requiring a network redesign or additional inline appliances. According to the company, customers can gain visibility in one hour, put protections in place in two weeks and achieve about 90% segmentation in 90 days. This can turn a compromised system into a dead end, even as AI makes it possible to explore networks and move between systems at machine speed. More than 400 enterprises use Zero Networks. The company says it has automated millions of policies, has a 99% customer retention rate and holds a 5.0 Gartner Peer Insights rating, with 100% of reviewers willing to recommend it. Learn more at www.zeronetworks.com.

View the original announcement on Business Wire: https://www.businesswire.com/news/home/20261021765411/en/

Contacts

Media Contact
Alex Daigle
[email protected]