Writing on web security

Rogue agent delegation is when a trusted agent spawns a sub agent that pushes a harmful action past a control that only…

An agent impersonation attack lets a rogue agent pose as a trusted one and issue orders others obey. Learn why it works…

Multi agent prompt injection is when one compromised agent poisons its peers. Learn how inter agent trust launders one injection and how…

A tool chaining attack composes safe AI agent tool calls into a harmful sequence that leaks data. Learn why per call guards…

Agent hijacking lets untrusted text seize an AI agent’s plan loop and swap its goal. See how it differs from a confused…

Audit logging for ai agents is your detection layer: what to record per turn, how to spot bad tool use, and how…

Use mcp tool pinning to fingerprint each tool definition at approval and block silent changes, stopping rug pulls, poisoning, and tool shadowing.

Egress filtering for ai agents is the containment layer that stops stolen data leaving: domain allowlists, stripped markdown images, and inspecting proxies.

Use prompt injection canary tokens to catch LLM data leaks: plant secret markers, watch every exit, and get hard proof the instant…

How llm guardrail models screen the input and output of a main LLM, where these filters catch attacks, and the honest limits…