Agentic AI Security

Hands-on security guidance for AI agents, MCP, and AI coding tools

Agentic AI security is the practice of defending autonomous AI agents, the MCP servers they call, and AI coding tools against attacks like prompt injection and credential leakage. Here, "agent" means an LLM-driven software agent that takes actions, not a real-estate, support, or sales agent.

OWASP LLM Top 10 Browse AI security tools Who writes this

The threat model

Four attack classes account for most agent compromises. Every guide and tool review on this site maps back to defending one of them.

Prompt injection

Untrusted text hijacks the model into following the attacker’s instructions instead of yours.

Excessive agency

An over-permissioned agent acts autonomously at machine speed once it is manipulated.

Tool & MCP abuse

Poisoned tool descriptions and over-broad scopes turn the tool layer into an attack path.

Credential leakage

Secrets in prompts, tool arguments, or agent memory leak and grant standing access.

Start where you need to

01 Tools

AI security tools

How to evaluate MCP scanners, coding-assistant guardrails, and agent runtime hardeners, with the criteria and tests that separate them.

02 Guides

DevSecOps guides for AI workloads

Hands-on guides for secret hygiene, scoped credentials, and least-privilege agent permissions.

03 OWASP

OWASP LLM Top 10 (2026)

Each LLM risk mapped, risk by risk, to the concrete defensive controls that mitigate it.

04 Checklists

AI agent hardening checklists

Numbered, verifiable controls that shrink the blast radius of autonomous agents.

Latest guides

AI agent guardrail tools compared: open source and commercial options An honest comparison of real LLM guardrail tools, NeMo Guardrails, Guardrails AI, LLM Guard, Lakera Guard, and Prompt Security, based on public documentation, with a verdict table and guidance on choosing one. Audit your AI agent setup: a hands-on self-audit walkthrough A practical self-audit for your AI agent: inventory tools, check permissions, gate high-impact actions, test prompt-injection exposure, handle secrets, log tool calls, and vet MCP servers. Includes a numbered walkthrough and a printable checklist. The confused deputy problem in AI agents and MCP A definition-dense explainer of the confused deputy problem for AI agents and MCP servers: how an agent holding broad credentials is tricked via prompt injection or a malicious tool into misusing its authority, and the capability-based, least-privilege controls that contain it. AI agent observability tools compared: the audit trail for a misbehaving agent An honest comparison of real AI agent observability and tracing tools, Langfuse, Helicone, Arize Phoenix, LangSmith, and OpenLLMetry, based on public documentation, with a security and audit-trail angle and a verdict table. Data exfiltration risks for AI agents How sensitive data leaks out through an autonomous AI agent: prompt-injection-driven exfiltration, tool calls to attacker endpoints, beaconing markdown images, over-broad data access, and log leakage, plus the egress controls that stop it. How to detect a compromised AI agent and respond Warning signs an AI agent is compromised, from unexpected tool calls and unknown egress to anomalous API spend and injected outputs, plus a numbered incident response procedure to isolate, revoke, preserve, scope, and remediate.

Map every LLM risk to a concrete control

The OWASP LLM Top 10 (2026), each risk paired with the defensive controls that mitigate it.

Read the OWASP LLM Top 10 How this site is written