Repolode
Submit
0x4D31

0x4D31/endpoint-ai-agent-abuse

A curated catalog of techniques and real-world cases showing how attackers abuse local AI agents through their configuration, state, tools, and inherited access.

View on GitHub
Stars
100
Forks
6
Open issues
0
Last commit
4 days ago

What is it?

EAA catalogs the ways attackers can abuse AI agents running on workstations, in development containers, and in CI runners. It describes how an attacker launches an agent, changes its instructions and configuration, collects its transcripts and state, and uses the access the agent holds to files, tools, and services.

Each technique in the catalog names the attacker action and its prerequisites, and is assigned a maturity level: feasible, demonstrated by public research or a malicious artifact, or observed in an incident or campaign report. Separate case records track whether a procedure was present, attempted, executed, or led to confirmed impact.

The project focuses on agents that run for a victim user or organization. It includes a machine-readable JSON catalog and schema, detection ideas with required telemetry, and a validator script for contributions.

Key features

  • 01Catalogs 21 techniques covering launch, configuration, state collection, MCP and tool abuse, and inherited authority.
  • 02Assigns each technique a maturity level (feasible, demonstrated, or observed).
  • 03Maps incidents, malicious artifacts, and research to specific techniques.
  • 04Provides detection ideas with required telemetry and their limitations.
  • 05Ships a machine-readable JSON catalog with schema and a dependency-free validator.

Use cases

  • Security teams building detection rules for AI agents running on developer workstations.
  • Red teams reviewing how an agent could be abused through its configuration or tool access.
  • Engineers assessing the risk of giving CI runners or dev containers agent access.
  • Researchers looking for a structured taxonomy and case list of agent abuse.

Quick start

  1. 1

    Open the main catalog in the repository

  2. 2

    Run the dependency-free validator from the repository root

    python3 scripts/validate.py

Things to know

  • The README says sources and technical claims still need separate review beyond the validator.

In collections

Open-source AI agent frameworks & appsMCP serversSecurity scanners, audits & threat intel tools

Topics

#ai-agents#security#threat-intelligence#mcp#prompt-injection#detection#red-team

FAQ

What is endpoint-ai-agent-abuse?
A curated catalog of techniques and real-world cases showing how attackers abuse local AI agents through their configuration, state, tools, and inherited access. EAA catalogs the ways attackers can abuse AI agents running on workstations, in development containers, and in CI runners. It describes how an attacker launches an agent, changes its instructions and configuration, collects its transcripts and state, and uses the access the agent holds to files, tools, and services.
Is endpoint-ai-agent-abuse free to use? What license does it have?
endpoint-ai-agent-abuse is open source under the CC0-1.0 license. Check the license terms before using it in commercial products.
Is endpoint-ai-agent-abuse still maintained?
Actively maintained. The last commit to endpoint-ai-agent-abuse was on Nov 3, 2026.
What can I use endpoint-ai-agent-abuse for?
Security teams building detection rules for AI agents running on developer workstations. Red teams reviewing how an agent could be abused through its configuration or tool access. Engineers assessing the risk of giving CI runners or dev containers agent access. Researchers looking for a structured taxonomy and case list of agent abuse.
How do I install endpoint-ai-agent-abuse?
Open the main catalog in the repository Run the dependency-free validator from the repository root: python3 scripts/validate.py
Where can I get endpoint-ai-agent-abuse?
The source code and installation instructions are on GitHub: https://github.com/0x4D31/endpoint-ai-agent-abuse

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects