LEARN
Securing AI agents, from the problem up
An agent with shell access inherits everything you can reach, and a tricked one looks exactly like a helpful one. These guides explain the mechanisms, what each layer of defence actually covers, and where every tool in this category stops. Each one ends with something you can run.
Vaults, Brokers, and Sandboxes: What Each One Stops, and What None of Them See
A vault, a broker and a sandbox each answer a different question about what an AI agent can do. This page says which question, where each one stops, and what is still open once all three are in place.
What Can a Hijacked AI Agent Actually Do?
Not a hypothetical. A concrete walk through what an agent running as you can reach, why its hostile actions look identical to its helpful ones, and the three questions that tell you your own exposure in about a minute.
Claude Code Security: What the Built-in Controls Do Not Cover
Claude Code ships permission prompts, allow and deny rules, and hooks. They decide whether an action runs. They do not look at what it carries. This page maps the gap, names the five ways it is exploited, and shows the one channel no tool on the market can see.
How AI Coding Agents Leak Secrets, and What Actually Stops It
An agent does not need a zero-day to leak a credential. It reads a file it was allowed to read, then makes a request it was allowed to make. Both halves are approved, and the combination is the breach. What each defence layer catches, and where each one stops.
MCP Security: The Trust You Grant Once and Never Check Again
An MCP server is code you gave your agent, plus text you gave your model. Both halves are dangerous, and approval happens once while use happens forever. The failure modes, and the one design rule that matters most.
OWASP Top 10 for Agentic Applications: A Working Checklist
The ten threats in plain language, with one question to ask about your own setup for each, and what a good answer looks like. At the end, how node9 grades against the list, row by row.