Security & oversight
AI workforce security: questions to ask before deployment
A vendor-neutral security checklist for AI employees and agents: identity, data, tool permissions, prompt injection, auditability, and incident response.
Short answer
Before deploying an AI workforce, assess the full system: data access, user identity, tool permissions, isolation, prompt-injection exposure, output validation, logging, retention, incident response, and human approval for consequential actions. Use vendor evidence and your own threat model; a checklist or certification claim alone does not establish that a use case is safe.
Identity and least privilege
Ask how users authenticate, how service identities are scoped, and whether access is reviewed and revoked promptly. Limit each agent to the minimum data and tools needed for its job. Separate read from write permissions, and require approval for high-impact or irreversible actions.[2]
Untrusted content and connected tools
Treat retrieved pages, uploaded files, emails, and tool responses as untrusted input. Ask how the product defends against prompt injection, data exfiltration, unsafe tool calls, and excessive agency. Test realistic adversarial cases in a controlled environment and verify that policy enforcement happens outside model instructions.[2][1]
Evidence, monitoring, and response
Review audit events, retention, redaction, administrator access, incident notification, recovery, and vendor subprocessors. Confirm you can investigate an output, identify its inputs and tool calls, suspend the workflow, and delete data when required. Establish an owner and escalation path before production use.[1]
Test and govern the actual use case
Document intended use, prohibited use, quality limits, fallback behavior, and review requirements. Test both normal and failure cases, repeat evaluations after changes, and track incidents and near misses. NIST and OWASP provide frameworks and risk prompts; adapt them to your environment and obtain qualified review for regulated use.[1][2]
Frequently asked questions
Can a prompt tell an AI agent not to misuse a tool?
Prompt instructions are not a security boundary. Enforce authorization, validation, and approval in deterministic application controls wherever possible.
Does a vendor checklist replace a security review?
No. Verify evidence against your architecture, data, contracts, and threat model, and involve your security and legal teams.
Sources and further reading
For informational purposes—not legal, financial, or security advice. Verify current sources and terms before making decisions.