AI agents can read data and take action, so they create a new attack surface. Traditional application security is necessary but not sufficient. Here is the checklist we use with banks, operators and government entities.
1. Know every agent you have
Agents are being created everywhere, in productivity suites, low-code tools and engineering teams. Start with discovery and an inventory. Platforms such as Zenity continuously find agents and copilots and flag risky configurations.
2. Give agents least privilege
- Scope every tool and API an agent can call.
- Use separate, auditable identities for agents.
- Require human approval for payments, data exports and configuration changes.
3. Defend against prompt injection
Agents read untrusted content such as emails, web pages and documents. Treat that content as data, never as instructions. Apply input and output guardrails and govern tool calls at runtime with controls such as SOAISEC's SentraGuard or Palo Alto Networks Prisma AIRS.
4. Red-team before launch
Attack your own agents with jailbreaks, injection and tool-abuse scenarios before customers do. SOAISEC Labs offers AI red teaming as a service.
5. Secure the supply chain
Agents run in containers built from open-source packages. Use hardened, signed images such as Chainguard and scan every pipeline.
6. Prove it to auditors
Map controls to ISO/IEC 42001 and your regulator's expectations and collect evidence continuously with a GRC platform such as LockThreat. See our AI Security service.