Leila Underwood

Agents demand runtime evidence collection that static audits cannot reconstruct.

Layering credentials, tools, and gateways prevents agents from crossing unintended boundaries.

Short-lived certificates and narrow permissions replace risky stored keys for autonomous agents.

Securing agent CI/CD requires analyzing both code and agent dependencies separately.

Custom SAST rules must model agent-specific attack patterns that generic libraries cannot detect.

Agent-written infrastructure fails in patterns human review never catches.

Agents exploit insecure infrastructure defaults faster than humans ever could.

OWASP split AI security into two frameworks because models now act, not just talk.

Stopping an agent's process leaves its credentials live and fully authorized to act.

Attackers hide malicious instructions in tool descriptions that agents blindly obey.

Classical authorization systems fail when agents delegate to other agents at machine speed.

Attackers are reusing classical malware tactics against AI agents as new targets for compromise.

Prompt injection and rogue agents demand security controls built for AI, not packets.

AI agents exploit standing credentials at machine speed.