Coding agents and agents with interpreter tools turn natural language into executable code. Prompt injection, hallucinated packages, or simple mistakes can lead to commands that install malware, delete data, or open reverse shells on the host the agent runs on.
Run agent-generated code only in isolated sandboxes with no host access, restricted network egress, resource limits, and short lifetimes. Validate commands against allowlists, never pass model output to eval or shell interpreters directly, and review generated code before it reaches production systems.
The same feature written the risky way, then with the fix applied.
How this risk has ranked in each OWASP release.
2026Latest
#05asi05-2026 Unexpected Code Execution (RCE)
