Zero Networks has launched a capability designed to stop compromised AI agents by enforcing network-level restrictions on where they can operate, even after they've been granted valid credentials. The company announced Monday its "Least Agency Enforcement" feature, which applies the Open Worldwide Application Security Project's emerging Least Agency principle using identity-based micro-segmentation to prevent agents from exceeding their intended scope. The offering aims to address a gap in current AI security approaches, which focus heavily on controlling what an agent is instructed to do rather than limiting the damage if those controls fail.
The announcement arrives as enterprises rapidly deploy AI agents without matching security safeguards. Research conducted by Zero Networks found that nearly 80% of organizations have already put internal AI agents into production, yet roughly two-thirds still lack governance policies for them. OWASP's Agentic Applications Top 10 project recently introduced the Least Agency principle, which recommends that organizations explicitly limit an AI agent's autonomy, tool usage, and decision-making authority to reduce risks including prompt injection, privilege abuse, and compromised agents.
"Most vendors are trying to control AI agents at the application layer," said Chris Boehm, Zero Networks' field CTO, according to the company's announcement. The company's approach differs by focusing not on what the agent was asked to do, but on what systems it can reach if manipulated, misconfigured, or simply incorrect. Benny Lakunishok, CEO and co-founder of Zero Networks, explained that the company is applying the same logic to AI agents that least privilege security did for people, except now it must be automated so that a fooled or misused agent "should hit a wall almost immediately, not wander around the network looking for something valuable."
The technology works by mapping which systems an agent identity should be permitted to access, then using host firewall enforcement to deny everything outside that set by default. Rather than competing with identity or privilege access management platforms that govern whether an AI receives credentials in the first place, Zero Networks positions its enforcement as addressing what happens after an authenticated session begins. The company argues that conventional identity tools generally don't control where an agent can move across the network once it has valid access. Zero Networks' system operates on both identity and network layers, restricting communication to authorized agents and routing sensitive protocols through multi-factor authentication as a backup against compromises. Boehm noted that sensitive paths get an MFA prompt on the protocols themselves, preventing a compromised agent identity from quietly using RDP, SMB, or WinRM to move sideways across systems.
The capability builds on Zero Networks' existing AI security portfolio, which includes AI Agents Control, AI Segmentation, AI SaaS Control, and protections for enterprise large language model deployments. Least Agency Enforcement is available immediately and will be demonstrated at Black Hat USA 2026, where the company plans to showcase its broader AI security platform. For organizations racing to adopt agentic AI, the pitch is simple: application-level controls focus on prompts and model outputs, but identity and network enforcement limit the business impact if those safeguards fail.