Runtime Safety
A security concept adapted from traditional software engineering, referring to real-time monitoring and intervention applied to model inference and tool calls during the active operation of an AI system. It verifies compliance with predefined policy constraints and acts to prevent unauthorized operations and harmful outputs from occurring during execution.
Timeline (last 90 days)
Runtime security should take actions as the audit unit, recording what tools the agent called, what resources it accessed, and what addresses it connected to, rather than recording only its text output.
Whether runtime security can protect Agents that can bypass the gateway depends on its definition: if it is merely policy code inside the gateway, it cannot protect them; if it is a defense-in-depth system covering credential management, network isolation, and token issuance, it can provide a fallback.
All Facts (2)
Runtime security should take actions as the audit unit, recording what tools the agent called, what resources it accessed, and what addresses it connected to, rather than recording only its text output.
50%UnverifiedWhether runtime security can protect Agents that can bypass the gateway depends on its definition: if it is merely policy code inside the gateway, it cannot protect them; if it is a defense-in-depth system covering credential management, network isolation, and token issuance, it can provide a fallback.
50%