Treat AI agents like interns: restrain permissions or risk runaway actions
AI agents can go from buying shoes to buying a car without guardrails
At the Snowflake Summit, panelists from Resolve AI, 1Password, and Tenable highlighted the escalating security risks of AI agents. Mayank Agarwal (Resolve AI) warned that agents without ironclad constraints could perform unintended actions—like buying a car when asked for shoes. Unlike predictable API integrations from two years ago, agent behavior is now unpredictable: they dynamically wire tools and access data paths on the fly. This opens the door to shadow AI and data exfiltration, as agents may read from one tool and write to another without oversight.
Further complicating governance is that agents look like humans or service accounts to systems, making attribution nearly impossible. Nancy Wang (1Password) stressed the need to know not just what an agent does, but whose authority it acts under and what it does with accessed data. Jason Merrick (Tenable) cited a client with 12 unauthorized OpenClaw instances accessing source code via Telegram. The solution: balance access with governance, enforce context and intent in agent design, and maintain deep human oversight rather than blanket blocking.
- Agents without constraints can trigger runaway actions—e.g., buying a car when asked for shoes.
- Agent behavior is now unpredictable; they wire APIs on the fly, unlike deterministic software of two years ago.
- Attribution is murky: agents can appear as humans or service accounts, making audits difficult.
Why It Matters
Organizations must implement tight permission controls and human oversight to prevent AI agents from causing data leaks or unauthorized actions.