AgentChaos: Open-Source Harness Tests Prompt Injection via MCP Tool Results
DiscussionHealthy802 · reddit · 2026-09-14
A developer released AgentChaos, an open-source harness for testing what happens when untrusted instructions arrive through MCP tool results. Across 356 valid trials on six models, it tested two concrete exploits: exfiltrating a planted credential and fetching a cloud instance-metadata endpoint.
Key finding: the dangerous part isn't just whether an agent calls a harmful tool — some agents read and accept injected instructions but only fail because the sandboxed workspace lacks a usable recipient. In a realistic workspace, agents that looked safe can complete the attack.
The author is soliciting feedback from MCP server/client builders on whether tool results should carry stronger provenance or be marked untrusted by default. Repo on GitHub; full write-up on Medium.
More from coding & agent
- 5 specialized inference engines in a month: why fork the ecosystem when vLLM/SGLang can hit near-SOL? — hsu_byron · 2026-09-14
- iOS AI dev workflow: AppKit + Figma import + Claude Opus 5 gives best UI fidelity — dotey · 2026-09-14
- Dev seeks blueprint for agent that triages incidents across PagerDuty, Datadog, GitLab and Slack — cruelcaricature · 2026-09-14
- Redpen CLI checks whether your coding agent's 'done' claim matches repo, test and build evidence — Scobleizer · 2026-09-14
- Infinite Bookshelf: open-source app generates a whole book from one prompt using Llama on Groq — Roger_M_Taylor · 2026-09-14
- Agent Harness explained: freeCodeCamp compares 10 tools from Claude Code to DeepSeek Harness — Roger_M_Taylor · 2026-09-14