Governance

MCP makes it easy for agents to reach your systems, which is exactly why its deployments need discipline. This guide covers the risks that recur in reviews, over-privileged servers, injection through tool results, unvetted servers, credential leakage, and the controls that address each.

MCP turns integrations into an inventory, but only if someone governs it. This guide gives enterprise IT a governance model: who owns what, the server lifecycle, the registry, classification policy, review gates, and what to report.

A computer-use agent has the access of the desktop it runs on and can be steered by what appears on screen. This guide sets out the controls that make it safe: isolation, scoped credentials, allowlists, injection defense, gates, and recording.

Most organizations have coding agents and no policy. This template gives engineering and security leaders a governance policy they can adopt in a week: tools, data, access, review, gates, audit, incidents, and who owns each.

Model deprecations arrive on the vendor's schedule and land on whoever is not prepared. This guide gives a governance practice for deprecation risk: inventory, calendar, evaluation readiness, routing, contracts, and a runbook that turns a forced migration into a scheduled one.

Every agent in production needs a way to be stopped in seconds by someone with the authority to do it. This guide designs the kill switch: where it acts, what it stops, how work reroutes, who may pull it, how it is tested, and how service is restored.

Permissions bound what an agent may do through its tools; sandboxing bounds what it can reach if it does something else. This guide sets out isolation for agent runtimes: environment, network, filesystem, secrets, resources, and the tests that prove containment.

An agent is only as safe as the tools it can call. This guide gives a permission design method: classify every action, separate the consequential ones, scope a principal to each, enforce gates at the gateway, and prove it with tests.

Agents are a new class of non-human identity: they act, they act for people, and they can be steered. This guide sets out how to issue, scope, delegate, review, and revoke agent identities using the identity platform you already run.

A penetration test scoped for a web app misses most of what breaks in an AI system. This guide defines the AI-specific scope: prompt and indirect injection, tool abuse, exfiltration, identity and credential paths, sandbox escape, retrieval poisoning, and supply chain, with what good evidence looks like.

The AI supply chain includes things software supply chain programs never covered: model weights, hosted model versions, prompt libraries, datasets, agent tools, and MCP servers from third parties. This guide sets out provenance, review, registry, signing, sandboxing, and monitoring for each.

The EU AI Act applies to US companies that place AI systems on the EU market or whose AI outputs are used in the EU. This guide covers extraterritorial scope, the risk tiers, high-risk system obligations for providers and deployers, general-purpose model duties, transparency rules, phased timelines, penalties, and how a US company prepares without duplicating its US program.

ISO 42001 is the first certifiable international standard for AI management systems, giving organizations a structured, auditable way to govern AI across its lifecycle. This guide explains the standard's structure and controls, its relationship to ISO 27001 and other frameworks, the certification process, who benefits from certifying, and how to prepare.

Data leaves the building through prompts, retrieved content, tool results, and logs. LLM data loss prevention puts policy at the gateway and redaction at every boundary, so what reaches a model, a log, or a third party is decided by rules rather than by whoever wrote the prompt.

A model's output is a proposal, not a result. Output validation checks it against schemas, business rules, groundedness, safety policy, and permission constraints before it reaches a user or triggers an action. This guide covers the validation layers, how to implement them deterministically and with judge models, failure handling, and testing.

The NIST AI Risk Management Framework is the voluntary US reference for managing AI risk, organized around four functions: Govern, Map, Measure, and Manage. This guide explains each function, the trustworthiness characteristics, the generative AI profile, how organizations adopt it in practice, and why regulators and statutes point to it.

An agent that holds an API key in its context window will eventually print it. Secrets management for agents means the model never sees a credential: the tool layer obtains short-lived, scoped credentials from a broker, traces are redacted, rotation is automatic, and leaks are detected.

SOC 2 does not have an AI section, and auditors still ask AI questions. This guide maps the Trust Services Criteria to model calls, prompts, retrieval, agents, and providers, lists the evidence that satisfies them, and shows how to prepare for a Type II observation period.

The Telephone Consumer Protection Act and its state counterparts shape what an AI voice agent may do, especially outbound. This guide covers consent, calling windows, do-not-call, identification, opt-out, recording, and automation disclosure, with the controls that turn rules into system behavior.

An AI bill of materials is the inventory of everything an AI system is made of: models and versions, datasets, prompts, libraries, agent tools, and providers, with provenance, licenses, and risk attributes. Regulators, auditors, and enterprise buyers increasingly ask for it. This guide explains what an AIBOM contains, how it relates to a software bill of materials, how to generate and maintain one, and how to use it in governance and procurement.

Red teaming simulates realistic adversaries against AI systems to find what evaluation and penetration tests miss. This guide covers setting objectives and rules of engagement, assembling a team, running campaigns across injection, manipulation, and misuse, scoring findings by impact, turning results into controls and tests, and cadence by risk tier.

The United States has no single AI law, but AI is regulated through existing consumer protection, civil rights, privacy, and sector laws, agency guidance, and a growing set of state statutes. This guide maps the layers, explains what they require of businesses, and sets out a preparation strategy that holds as the landscape shifts.

AI systems handle more secrets than most applications: provider API keys, credentials for every tool an agent calls, and whatever sensitive strings pass through prompts. This guide covers gateway custody of provider keys, runtime injection of tool credentials, keeping secrets out of prompts, context, memory, and logs, rotation, and detection of exposure.

Third-party risk programs were built for vendors that host data and run services. AI adds model providers whose behavior changes, AI vendors that process sensitive data through subprocessors, and AI features appearing inside software already procured. This guide covers inventory, tiering, due diligence, contracts, ongoing monitoring, and concentration risk.