Skip to content
AS
13 / AI & Data2026

LLM Guardrail Breaker

An adversarial testing framework for AI agents, MCP servers, and RAG workflows covering injection, tool poisoning, SSRF, and data exfiltration.

PythonMCPRAGSandboxingAdversarial ML
Case file
Role
Solo build — research, engineering, documentation, release
Status
Planned · frontier build
Year
2026
Position
13 of 13
Case 13 / llm-guardrail-breakerArchitecture · Evidence · Signal

01 — Context

The problem worth solving.

Agent security fails at the boundary between model output, tools, data retrieval, and user privileges.

02 — Approach

How it was built.

  1. 01

    Automate indirect prompt injection and cross-tenant retrieval tests.

  2. 02

    Test SSRF through agent tools and malicious MCP schemas.

  3. 03

    Produce risk reports plus scoped tool wrappers and deterministic controls.

03 — Architecture

The shape of the system.

01Target

Agent, MCP server, retrieval layer, tools, and privilege context.

02Attacks

Injection, poisoning, SSRF, and data-exfiltration scenarios.

03Controls

Input/output firewalls, sandboxing, and least-privilege tool wrappers.

04 — Evidence

Receipts, not adjectives.

MCPtool surface
RAGdata plane
CVSSrisk format

05 — Hiring signal

“Treats agent security as an end-to-end systems problem instead of a prompt-only problem.”

Ships working systemsDocuments the methodExplains the trade-offsKeeps evidence attached

06 — What's next

Still being built.

Define the attack corpus and implement the first tool-poisoning harness.