A new paper from University of Chicago researchers says securing AI agents requires treating them like networked systems rather than trusting the agents themselves. Current defenses mostly put the burden on the agent to spot threats and enforce privacy rules. That fails because the language models driving agents are unpredictable and open to attacks such as prompt injection. The authors borrow long-standing networking ideas: centralized control with distributed enforcement, capability-based access, and least-privilege zero-trust rules. These give strong, predictable guarantees in traditional networks. They are not enough alone for agents, though. Whether an action is safe often depends on context that fixed rules cannot capture. It sketches a reference architecture where each agent sits behind a sidecar that mediates every request, message, and external call, while a separate control plane stays outside the agents’ reach.
A new paper from University of Chicago researchers says securing AI agents requires treating them like networked systems rather than trusting the agents themselves. Current defenses mostly put the burden on the agent to spot threats and enforce privacy rules. That fails because the language models driving agents are unpredictable and open to attacks such as prompt injection. The authors borrow long-standing networking ideas: centralized control with distributed enforcement, capability-based access, and least-privilege zero-trust rules. These give strong, predictable guarantees in traditional networks. They are not enough alone for agents, though. Whether an action is safe often depends on context that fixed rules cannot capture. It sketches a reference architecture where each agent sits behind a sidecar that mediates every request, message, and external call, while a separate control plane stays outside the agents’ reach.