AI agents change IT security because they can send mail, alter files, and invoke APIs. The current risk landscape centres on shadow AI, shadow agents, prompt injection, excessive permissions, exposed secrets, and incomplete audit trails.
Controls that matter
Give every agent a distinct identity with least-privilege and preferably time-limited access. Default-deny outbound network routes. Require approval for consequential actions, provide a kill switch, and keep audit logs outside the agent’s write permissions.
Inventory private AI accounts and unofficial workflows. Separate untrusted retrieved content from tool instructions and never place API keys in prompts. Local AI reduces provider data flows but does not prevent prompt injection or misuse of powerful tools.
Connect sovereign AI, GDPR-compatible agents, and role-based AI literacy.
Secure AI agents and shadow AI
We inventory data paths and agents, constrain identities and tools, and implement approvals, audit logs, and kill switches.