AI Security Guides & Reviews

Threat models and defenses for LLM applications and autonomous agents, grounded in current research.

Home › AI Security

Giving a language model tools, memory, and network access changes what a successful attack costs. A jailbroken chatbot produces bad text; a compromised agent takes actions. This category covers that shift — the OWASP Top 10 for LLM Applications, indirect prompt injection as the vector that turns retrieved content into instructions, Model Context Protocol and tool-poisoning risk at the protocol layer, and the broader agent threat model including the combination of private data, untrusted input, and an available exfiltration path.

The coverage tracks published research and vendor disclosures rather than speculation, and it is explicit about what remains unsolved. Prompt injection has no known complete fix; what exists is a set of layered mitigations — least-agency design, provenance-aware prompting, injection classifiers, human confirmation for irreversible actions — each closing part of the surface. Where a defense is partial or evadable, the guides say so instead of presenting it as a solution.

Published in AI Security (5)