AI Sec Watch: A Security Intelligence Platform for AI Systems

Luu, T.J.

Securing AI agents: Key controls and best practices

infonewsLLM-Specific

securitysafety

Source: CSO OnlineSeptember 8, 2026

Summary

AI agents pose unique security risks because they operate at machine speed, can chain multiple actions together into unauthorized outcomes, and inherit the credentials and access of employees, but existing security controls designed for humans are inadequate to stop them. Security experts warn that traditional approaches like system prompts (instructions given to an AI model) are not hard blockers, and organizations need technical controls outside the agent's control to detect when agents cross unauthorized boundaries, immediately revoke their credentials, and roll back their actions.

Solution / Mitigation

According to the source, organizations should: (1) deny direct internet access by default and route requests through proxies that enforce domain and operation allowlists (approved lists of allowed domains and actions), (2) separate read and write capabilities so agents cannot both access and modify data equally, (3) require explicit approval for high-risk actions such as deletion, privilege changes, and data exports, (4) limit an agent's available functions, permissions, and autonomy, and (5) enforce authorization in downstream systems (the actual tools and services the agent uses) rather than relying on the AI model to decide what is permitted. One expert example mentioned was using a 'scope-level hook on every command-line tool call' that checks each command against an approved target list.