Topics in agentic AI security
Plain, source-cited reference pages on the threats, terms, and incidents that define agentic AI security. Each page is written answer-first, cites primary sources, and links to the underlying research. This library is being built out; the pages below are live.
Published
Agentic AI Security
Securing autonomous AI agents: why it is a distinct discipline, the main threats, the three-layer defense model, least agency, and the standards. The hub the rest of this cluster links back to.
MCPMCP Security
The Model Context Protocol attack surface: tool poisoning, rug pulls, cross-server shadowing, and why the connect-time-vs-runtime gap is the root issue.
Prompt injectionIndirect Prompt Injection
Hidden instructions in the content an agent reads — the mechanism behind EchoLeak — and why identity controls alone cannot stop it.
MCP SecurityTool Poisoning
Malicious instructions hidden in a tool's description that steer an agent into acting for an attacker.
In progress
The following reference pages are being written to the same standard and will be published into this section. Until each ships, the research papers and the glossary cover the same ground.
- Prompt Injection — The foundational attack class, direct and indirect.
- Runtime Agent Security — Inspecting and enforcing on the action, at the point where an agent acts.
- The OWASP Agentic Top 10 — A reference walkthrough of ASI01-ASI10.
- Non-Human Identity — Securing an agent as a scoped, least-privilege identity.
- EchoLeak (CVE-2025-32711) — A factual, cited breakdown of the first zero-click prompt-injection chain.
How this section is built. Every topic page follows one template: an exact-term H1, a bottom-line-up-front answer, an extractable “key facts” list, question-headed sections answered answer-first, inline primary-source citations, a brief factual note on how Crawdad addresses the issue, and cross-links to related pages and the relevant white paper. Each carries TechArticle, FAQPage, and DefinedTerm structured data. See Tool Poisoning for the pattern.