Securing AI Agents Requires Least Privilege and Continuous Access Controls

Summary
The article argues that AI agents inherit risks from over-privileged non-human identities and recommends inventory, least privilege, just-in-time access, short-lived credentials, continuous validation, and session-level accountability.
Key points
- AI agents may use long-lived API keys, tokens, certificates, or other credentials with broad permissions and weak ownership or offboarding.
- The article warns that prompt injection can steer an agent with real permissions into unintended actions.
- It recommends inventorying each non-human identity, assigning an owner and purpose, and setting an expiry.
- Use least privilege and just-in-time access, with short-lived, scoped credentials where possible; eliminate or rotate static secrets and vault those that cannot be rotated.
- Give each agent a distinct identity, continuously validate its trust, and maintain session-level visibility and the ability to revoke its authority.
- The article cites OWASP's 2025 Non-Human Identities Top 10, which ranks improper offboarding among the leading risks, followed by secret leakage.
- CREST's AI Charter and the NCSC/DSIT Cyber Shield blueprint are cited as recent initiatives emphasizing accountability, identity, and trust in AI-enabled cybersecurity.
Article Details
- Defense Focus
- Reduce risk from AI agents and other non-human identities by controlling their privileges, credentials, and sessions.
- Detection Methods
- Inventory non-human identities, their owners, permissions, credentials, and expiration dates.
- Continuously validate agent identity and trust against behavior.
- Monitor agent sessions to identify which identity acted, under whose authority, and against what resources.
- Review repositories and configuration files for static or hardcoded secrets.
- Data Sources
- Non-human identity and service-account inventories
- Credential stores, repositories, and configuration files
- Authentication records
- Agent behavior and session activity
- Defensive Actions
- Apply least privilege and issue scoped credentials just in time, with short expiration windows.
- Eliminate long-lived secrets where possible; rotate or vault secrets that cannot be eliminated.
- Assign each agent a distinct, attributable identity with an owner, purpose, and expiry.
- Continuously verify identity and trust rather than relying on implicit trust or a one-time login.
- Maintain session-level accountability and a kill switch to revoke an agent’s authority quickly.
MITRE ATT&CK
T1078 · Valid AccountsThe article discusses attackers using stolen credentials to log in, including credentials associated with non-human identities; it presents this as a general threat pattern, not a specific incident.T1552.001 · Credentials In FilesThe article describes attackers hunting for credentials that are hardcoded in repositories or stored in configuration files.
