Overview

Make agents safe to ship

An AI agent that can act can also cause real harm. One wrong move can leak data, break a system, or ship bad code. This guide shows you how to stop it: build, check, test, and release agents you can actually trust, held to the standards your security team will judge you by. Pick a section and dig in.

How the guide fits together

Secure by design · Field guide

AI Agent Anti-Patterns

The recurring design mistakes that turn capable agents into liabilities — wildcard tool exposure, standing credentials, prompt-injection paths — each with the attack, the fix, and how to verify it. Twenty-five named failure modes, grouped so you can find the one that matches your architecture.

What is inside

• Identity & access

  Wildcard tool exposure, standing credentials, shared identity runtime.

• Prompt injection

  Unauthenticated tool channels, conflated context, poisoned knowledge bases.

• Supply chain

  Live-fetch dependencies and mutable reference trust.

• Governance

  Unsupervised perimeter, unattributable action, boundaryless data egress.

Supply-chain trust · Field guide

Vetting the Code You Bring In

Every system is mostly other people’s code — packages, actions, images, models, tools. This series is the discipline most teams skip: deciding whether outside code is safe to trust, before you trust it.

What is inside

• Reading the risk of an unknown repository

  Read the provenance, verify independently, inspect the danger surface, and handle it safely.

• Vetting an open-source package

  Decide whether a package is safe before you add it.

• Trusting a third-party CI/CD action

  Judge an outside build step before it runs in your pipeline.

• Trusting a container base image

  Check a base image’s provenance and contents before you ship on it.

Secure delivery · Field guide

Releasing AI Agents to Production

A practical release pipeline for agentic systems — pinned dependencies, ephemeral CI runners, signed images, chain of custody, and rollback — so what you tested is exactly what ships.

What is inside

• Delivery models & chain of custody

  How the agent reaches the customer, and who is accountable along the way.

• The pin-file contract

  Signed images and a pinned manifest so the build is reproducible.

• Ephemeral CI runners & hygiene

  Disposable build environments and a pre-flight hygiene checklist.

• Cadence & rollback

  Releasing on a predictable rhythm with a safe way back.

Post-quantum readiness · Field guide

Getting Quantum-Safe

From first principles to an enterprise migration plan: what a large quantum computer will break — RSA, ECC, TLS, PKI — the finalised NIST standards that replace today’s cryptography, and a structured post-quantum readiness assessment, ending with what it all means for AI.

What is inside

• Foundations

  Secrets, keys, encryption, and quantum computers — in plain language, no maths assumed.

• The threat

  Harvest-now-decrypt-later, and exactly what breaks — RSA, ECC, PKI, TLS.

• The new rulebook

  The finalised NIST standards that replace them: ML-KEM, ML-DSA and SLH-DSA.

• The readiness assessment

  A five-phase method with a CBOM, risk classification, and a 0–5 maturity model.

Self-test toolkit

Self-Test for AI Agent Anti-Patterns

Copy-paste red-team prompts and a fix-and-retest loop that probe your agent for every anti-pattern in the catalogue — run it any time, before any release. The actionable companion to the anti-patterns field guide.

What is inside

• The prompt pack

  Twenty-five targeted security checks you paste against your own agent.

• The red-team prompt

  A single adversarial prompt to stress the whole agent at once.

• Fix-list and retest loop

  Turn findings into fixes, then prove the fix held.

• Ten-minute pre-release audit

  A fast, repeatable gate to run before every release.