Defense · 8 Chapters

Security

The privacy and trust surface of building with agents — what data leaves, what the model retains, what a tool can do on your behalf, and how an attacker turns your own context window against you. Not generic appsec. Not model alignment. The boundary between your secrets and a stateless model you do not control.

The Eight Chapters

Front to back: the risks first, then the threat model, then the defenses. It cuts across the other three pillars, not beside them.

The Real Risks

What actually leaves your machine, what the model retains, and what an agent can do on your behalf — the honest threat landscape.

AvailableRead

Data & PII in Context

What enters the context window, provider retention and training exposure, data minimization, and redaction that survives the loop.

AvailableRead

A Threat Model for AI

The attack surface of an agent: what is genuinely new versus classic appsec, and the four trust boundaries — model, context, tools, output.

AvailableRead

Prompt Injection

Direct and indirect injection, tool-result poisoning, the limits of the instruction hierarchy, and why there is no single fix.

AvailableRead

Secrets & Credentials

Keys in prompts, env, and logs; the never-in-context rule; scoping and short-lived credentials; agent credential isolation and secret scanning.

AvailableRead

Permission Architecture

Least-privilege tools, allow/deny, sandboxing, blast radius and reversibility tiers, human-in-the-loop gates, and SSRF via the agent fetch.

AvailableRead

Output & Exfiltration

Exfiltration via tool calls, links, and markdown image beacons; the danger of rendering untrusted model output as live HTML.

AvailableRead

Supply Chain & Audit

MCP, plugin, and model provenance; dependency and lockfile hardening; what to log and what never to log; traceability for autonomous actions.

AvailableRead

Where This Connects

Security is not a fourth silo — it is the trust surface cutting across the other three. These patterns, toolkit deep-dives, and harness chapters are where it grounds out.

Negative SpacePattern Scope FencePattern Safety NetPattern HooksToolkit Settings & PermissionsToolkit System Prompt ArchitectureHarness