Agentic AI does not fail because teams cannot build agents. It fails because teams cannot govern what agents are allowed to see, decide, and do.
Bounded AI Autonomy is the practitioner field guide to AI agent security and AgentSecOps for engineers, architects, and security leaders putting agentic AI into production inside organisations where failure has real consequences: financial services, healthcare, regulated industries, and any environment where a data leak, a confused-deputy attack, or a runaway agent ends up in a board pack.
The book covers a seven-layer reference architecture for production agents, with explicit trust boundaries at every layer. It treats MCP as both a useful protocol and an attack surface, with chapters on supply-chain hygiene, tool poisoning defences, manifest signing, and sandboxed servers. It introduces the Confused Deputy problem and the OAuth 2.1 token exchange (RFC 8693) pattern that fixes it, with working Python code, then layers asymmetric dual-LLM guardrails, capability matrices, Agent Contracts, and policy-as-code in Rego and Cedar on top.
Secure RAG gets its own part, covering permissions-aware retrieval, classification at ingestion, indirect prompt injection sanitisation, and citation validation. The operations chapters treat evals as the new test suite, observability that supports full trace replay, agent reliability engineering for non-deterministic systems, and a unified AgentSecOps architecture with build-time gates and runtime enforcement that close the loop from incident back to gate.
Twenty-one chapters. Forty-two technical diagrams. Code examples in Python, Rego, Cedar, YAML, and shell. Opinionated, code-heavy, and calibrated for environments where regulators ask hard questions. The book treats every AI agent as a new, highly privileged employee — one that needs identity, audit, restriction, and a clear performance review just like a human colleague.
If you have moved past AI demos and need to ship, defend, and operate AI agents in front of real customers, regulators, and auditors, this book is for you.
