Notes on security for AI agents
Agent identity, permissions, audit trails, and engineering for secure autonomous systems.
Why AI Agents Must Not Inherit Your Credentials
Service account tokens let autonomous agents inherit accumulated access. That blurs scope, rotation, and attribution. Separate agent identity makes permissions narrower and actions traceable.
Read article
Scoped Access Versus Broad API Tokens for Agentic Workflows
A long-lived API token can simplify multi-service orchestration. Risk emerges when an agent uses that broad access in ways nobody intended.
Read article
Audit Trails for Autonomous Agents
When a postmortem credits a person for an action an agent performed, the audit trail is more than incomplete: it misdirects the response team.
Read article
Practical Primer: Identity Verification for AI Agents
To set an agent's permissions, first establish which agent sent the request. Verified identity gives the authorization layer a trustworthy basis for its decision.
Read article
Designing Permission Schemas for Multi-Agent Pipelines
With multiple agents, permissions compound. A schema that suits one agent can fail quickly once agents pass context between one another.
Read article
AI Agent Authorization: RBAC vs ABAC
Role-based access suits people with stable roles. Agents change context as they work, giving attribute-based models a clear advantage.
Read article
Where Platform Teams Miss Agent Security
Most platform teams view agent security as service account security, but subtle differences produce very different failure modes in practice.
Read article
Scaling Credential Rotation for Autonomous Systems
Rotating one engineer's credential is routine. Rotating credentials for 40 agents across 12 services at once is a fundamentally different engineering challenge.
Read article
Hidden Risks of AI Agents Using Dev Team Tokens
A shared developer token in your CI environment may already run an AI agent. Learn what that means for your blast radius when something goes wrong.
Read article
Least-Privilege for Autonomous AI Agents
Least privilege is established practice. Applying it to subjects that can reason about and request their own permissions is newer, with failure modes that are not obvious.
Read article
Scoping Down AI Agent Permissions
Begin with a deliberate permission hypothesis, observe the agent in operation, and narrow its scope step by step as you learn what it actually needs.
Read article
Governing Your First AI Agent Deployment
Defining AI agent governance is straightforward in theory, but much harder before launch, when there is no real activity data to show what the agent actually does.
Read article