White papers
Plain, source-cited research on securing autonomous AI agents. Each paper is available as an indexable web page and as a PDF. The voice is deliberately non-promotional; the figures are drawn from public research and, where they are Crawdad's own, are reproducible.
The Agentic Attack Surface
What changes when your AI starts to act — the OWASP Agentic Top 10, the incident record, MCP tool poisoning, and why the model separates identity from runtime.
Reference architectureA Reference Architecture for Runtime Agent Security
Where the runtime layer sits and how to build one that is both effective and trustworthy: on the wire, on the device, legible and testable, provable.
Field briefingThe State of Agentic AI Security
A working map of the landscape: the taxonomy, the incidents, the MCP surface, the three-layer defense model, readiness data, and the regulatory calendar.
The benchmark corpus
Alongside the papers, Crawdad publishes an open, reproducible benchmark of attacks targeting modern LLM agents. Anyone can clone it, point it at any HTTP-addressable classifier, and check the numbers. It is licensed CC BY 4.0.
| Part | Count | What it is |
|---|---|---|
| Attacks | 497 | Across 13 categories (395 visible + 102 holdout) — prompt injection, credential exfiltration, indirect injection via tool output and RAG, tool abuse, supply-chain manipulation, memory poisoning, agent-to-agent compromise, and more. |
| Benign samples | 1,172 | Across 4 negative categories (developer messages, conversational, security discussion, hand-curated near-misses) — for measuring the false-positive rate honestly. |
| Runner | any tool | Tool-agnostic: works with any classifier that accepts POST {"text": "…"} and returns blocked vs. allowed. |
| License | CC BY 4.0 | Use it for your product, paper, or class. |
A detection number on its own is a vanity metric — blocking every message scores 100%. The corpus ships negatives so a tool's false-positive behavior is measured alongside its detection. Run it yourself and compare.
Topics
The Topics section turns the research above into answer-first reference pages for the terms people and AI engines search. Start with Tool Poisoning.