Agent Attack Paths Measured: Browser Content Leaks 11/12, MCP Config 8/8, Shell 0/14
DaimoNNN · reddit · 2026-09-21
A hands-on test of three entry paths into AI agents quantifies prompt-injection leak rates: browser content leaked in 11 of 12 runs, project-scoped MCP config leaked in 8 of 8, while shell commands leaked 0 of 14. The takeaway: the wider an agent's context sources, the wider its injection surface — MCP config files and web content are the weakest links. Full writeup at efecakici.com.tr/blog/agent-lab-three-paths/
More from coding & agent
- Fine-tune your own model for free: Claude Code writes the code, Kaggle T4s do the training — ojasvi_yadav · 2026-09-21
- Google open-sources Agent Substrate, a K8s agent runtime that suspends and resumes per tool call — blaizedsouza · 2026-09-21
- Tencent open-sources T-Mem, a memory architecture that anticipates instead of archiving — blaizedsouza · 2026-09-21
- Ingesting 100k papers into an LLM wiki: user seeks editorial policy for AI knowledge synthesis — gintrux · 2026-09-21
- Chat On Steroids: open-source Codex-style agent that runs on your ChatGPT subscription quota — CurieuxExplorer · 2026-09-21
- Vibe coding as the next abstraction layer above the compiler — binarybits · 2026-09-21