Code Execution Defense

Explanation

Code execution defense checks that an AI system — especially an agentic one that can run code, call tools or execute commands — cannot be manipulated into running attacker-supplied or unintended code. In BFSI this is among the most dangerous risks because code execution can be chained into data exfiltration, fund movement or full system compromise. It is measured by a Code Execution Vulnerability Rate; no calculation method is provided in the control data, so implement it as the proportion of code-execution attack probes that succeed (vulnerable executions divided by total probes) and document the exact definition you adopt. To implement and operate it, run any model-generated or tool-invoked code in a tightly sandboxed, least-privilege, network-restricted environment with strict input validation and an allowlist of permitted operations, place this enforcement directly in the agent's tool-execution path, and red-team it with injection and escape probes both before release and continuously in production; log every execution request, the sandbox decision, and any escape attempt as evidence. The threshold is ≤ 1% (target 0) with an alert on any breach: ideally no probe achieves unauthorised execution, and any confirmed vulnerability is an immediate-priority event that should disable the affected tool or agent, contain the environment, and trigger patching and investigation before reactivation.