Direct answers to the questions teams hit putting AI agents into production: what the terms mean, how the pieces differ, and where the risk actually sits.
By entry type
Agents gain capabilities through MCP servers, skills, plugins, extensions, packages and SaaS connectors. Each behaves differently at install, at execution and at review. Here's the taxonomy, and why allow-listing alone doesn't govern it.
How organizations attribute, predict, and control what autonomous agents spend, and why cost and security are the same instrumentation problem.
What actually turns a model into an agent, and why security controls have to target that layer rather than the model itself.
Where hooks sit in an agent's execution, what they can and can't protect against, and how they compare across major platforms.
Agent skills carry a specific security property: parts of a skill can execute before the model reasons about whether to use it. This entry explains how that happens, why refusal training and gateways both miss it, and what to check.
An AI agent audit trail is a durable record of what an agent did, under whose credentials, and in what context. This entry covers why agent logs are weaker evidence than most teams assume, what immutability actually means, and what auditors ask for.
What agent behavior means, how it differs from architecture and configuration, and why agent risk accumulates across sequences rather than appearing in single actions.
Three terms that are often used interchangeably, for three jobs that depend on each other. What each one means, the fields an inventory needs before it can support a decision, and how to build one across cloud, code, endpoint and browser.
The EU AI Act, NIST AI RMF, ISO 42001 and the OWASP Agentic AI Top 10 applied to autonomous agents: what they have in common, why agents are harder to evidence, and why compliance is downstream of discovery.
The continuous practice of connecting what an agent is configured to do, what it can actually reach, and what it's actually doing, and how that differs from identity governance and point-in-time AI-SPM scans.
Cloud security secures the environment an agent runs in – not the agent's configuration, tools or decisions. What CSPM, CWPP and CIEM cover, and what they miss.
NDR inspects the traffic an agent generates, but the riskiest part of what an agent does never crosses the wire. What network detection covers for AI agents, and what it can't.
What AIDR means, how it differs from AI-SPM and guardrails, and the question that separates real AIDR from monitoring with a new name.
Deterministic controls produce the same outcome every time for the same input, while probabilistic controls use a model to judge each case. This entry explains why the distinction decides what you can audit, certify and operate at scale.
A gateway governs the agent traffic routed through it, while much of what an agent does never reaches one. This entry follows a single coding agent through a working session to show what each approach sees, and how the two fit together.
EDR protects the device an agent runs on. It doesn't see the agent's configuration, tool connections, or decisions. This entry sets out what EDR covers for AI agents, what it structurally can't, and where agent security sits alongside it.
An MCP gateway sits between agents and MCP servers, applying policy to the traffic that passes through it. This entry covers what it does well, the tool types it structurally cannot see, and the architectural question worth asking before you rely on one.
An MCP server exposes tools to an agent over a standard protocol. Its tool definitions are untrusted input, and approval doesn't survive the server changing.
MCP servers and agent skills solve different problems and are often confused as competing standards. This entry sets out the architectural difference between them and the security distinction most comparisons miss.
Non-human identity governs what an agent is permitted to access and under whose credentials. It doesn't establish what the agent did or who answers for it. This entry sets out where the two differ and why both are needed.
Prompt injection is when content an AI agent reads is treated as instruction. This entry covers how direct, indirect and stored injection differ, why it can't be patched like SQL injection, and what actually reduces the risk.
What shadow AI is, how it differs from agent sprawl, why surveys and gateways miss it, and how detection actually works.
Six criteria that decide whether a discovery tool finds your whole agent estate or only the agents you already knew about, with the question to ask a vendor for each and a way to test them in a trial.
No entries match that filter.
