OWASP Top 10 for Agentic AI: How to Protect Your AI Agents

#security #ai #python #opensource

OWASP released their Top 10 for Agentic AI Systems. Here is what matters and how to address each risk.

The Top Risks

Excessive Agency - agents doing more than intended
Supply Chain Vulnerabilities - compromised tools and plugins
Insecure Output Handling - agents producing unsafe outputs
Insufficient Logging - no audit trail of agent actions
Broken Access Control - agents accessing unauthorized resources

Addressing These with Governance

Most of these risks come down to: no visibility, no control, no proof.

Policy enforcement (risks 1, 5)

from asqav import Asqav
client = Asqav(api_key="sk_...")

# Block excessive actions
client.create_policy(
    name="limit-external-calls",
    action_pattern="api:external:*",
    action="block_and_alert",
    severity="high"
)

Audit trails (risk 4)

agent = client.create_agent(name="my-agent")
sig = agent.sign("data:read:users", {"query": "active"})
# Every action now has a quantum-safe cryptographic record

Multi-party authorization (risks 1, 5)

Critical actions require human approval before execution.

CI/CD Scanning

The asqav-compliance scanner checks your codebase for these risks:

- uses: jagmarques/asqav-compliance@v1
  with:
    standard: eu-ai-act

DEV Community