Seven Security Fundamentals for Limiting AI Agent Risks

Summary
Erich Winkler argues that AI agent risks often stem from longstanding security gaps. He outlines seven defenses, including least privilege, trust boundaries, data classification, secret management, supply-chain checks, and out-of-band verification.
Key points
- Limit each agent’s permissions to its task; use scoped tokens and separate identities.
- Keep agents exposed to untrusted content from also accessing private data and communicating externally.
- Treat model output as untrusted input; use parameterized queries, output encoding, and allowlisted commands.
- Classify organizational data and set clear rules for which tiers may be used with AI tools.
- Keep secrets in a vault rather than prompts or configuration files, and rotate exposed credentials.
- Verify models, packages, and MCP servers; pin versions, check hashes, and track components in an SBOM.
- Use callbacks to known numbers and dual approval for high-value payments or access changes.
Article Details
- Defense Focus
- Limit the damage an AI agent can cause by applying established security fundamentals, including least privilege, trust boundaries, data classification, secrets management, supply-chain integrity, and independent verification.
- Detection Methods
- Review agent permissions and assess what a compromised agent could access, change, or send.
- Identify agents that combine access to private data, exposure to untrusted content, and the ability to communicate externally.
- Check prompts, MCP configuration files, and repositories for exposed secrets.
- Verify model, package, and MCP server versions and hashes; maintain an SBOM that includes models and plugins.
- Data Sources
- Agent permissions and access
- Prompts and MCP configuration files
- Repositories
- Models, packages, and MCP servers
- SBOMs covering models and plugins
- Defensive Actions
- Give agents only the permissions needed for their tasks; use read-only access by default, scoped tokens, and a separate identity for each agent.
- Separate agents that read untrusted content from private data and outbound communication channels.
- Treat model output as untrusted input; use parameterized queries, output encoding, and allowlisted commands before passing output to other systems.
- Classify organizational data and define which data may be used with AI tools at each classification tier.
- Store secrets in a vault and inject them at runtime rather than placing them in prompts or configuration files; rotate secrets that have appeared in prompts or repository configuration.
- Pin versions, check hashes, prefer safe model formats, and use an internal registry or allowlist for packages, models, and MCP servers.
- Use an SBOM that includes models and plugins.
- Verify payment or access changes through a callback to a known number, dual approval for high-value transactions, or a pre-agreed executive verification phrase.
