Gen Outlines Security Measures for AI Agents

Gen says AI agents can be manipulated by malicious emails, websites and plugins into running commands, sending data outside their environments or deleting files. In a controlled experiment, researchers at Gen Threat Labs triggered those actions by presenting an agent with untrusted content. They said no software vulnerability was required.

The company is developing the Agent Trust Hub (ATH), a set of tools for assessing agent capabilities and controlling their activity. Gen says it includes skill scanning, a curated marketplace and checks agents can use to assess skills. Earlier this year, Gen researchers found that 12% of skills on ClawHub were malicious at the time. Gen also announced a partnership with Vercel to add verification and risk ratings to skills.sh.

ATH also includes VPN for Agents, which Gen describes as an AI-focused VPN powered by Norton VPN infrastructure. Its multi-tunnel design lets multiple agents use separate protected connections, including across different countries.

For AI coding assistants, Gen’s Sage can intercept shell commands, file operations and web requests before they run. Gen says it evaluates actions using URL reputation, threat heuristics, prompt-injection detection and supply-chain checks, then can allow an action, ask the user or deny it. Sage supports Claude Code, Cursor, VS Code, OpenClaw and OpenCode, and is open source.

Gen says Norton AI Agent Protection in Norton 360 monitors supported agent activity. Coverage depends on whether an integration exposes an action and lets the security engine stop it. Gen also published AARTS, an open specification for platforms to expose security events and enforce security decisions.