Insecure AI Agent Permissions
Critical · AI Security
OWASP Top 10:2025 (current)
Not directly mapped to a Top 10:2025 category — see CWE / CVE below.
CWE
See references
CVE / class examples
- Search NVD/CVE for product-specific instances of this CWE
All refs
OWASP Agentic
Definition
Agents granted excessive API/shell access.
Root cause
Auto-approve all tool calls.
Affected components
- Components in the “AI Security” class
- Application / API handlers
- Supporting libraries
Impact
Mass data deletion, fraud, lateral movement.
Severity
Critical · CVSS 9.3
Safe lab reproduction
lab-agent-perms
Detection methodology
Audit agent action logs.
Evidence examples
- Lab reproduction notes: lab-agent-perms
- Request/response excerpts (redacted)
- Detection signal: Audit agent action logs
Remediation
Per-tool scopes; step-up auth; deny by default.
Secure coding examples
Treat agents as privileged users.
Related CYNARKS lessons & labs
Tools: giskard
Practise related skills in Labs or deepen on Academy library.