Undergrad designs prompt-injection attack/defense benchmark for coding agents

Character_Cable997 · reddit · 2026-11-17

A second-year undergrad shared a semester project on Reddit: build a basic AI coding agent that reads GitHub issues, inspects repo files, edits code, and proposes fixes — then probe its vulnerability to indirect prompt injection by planting malicious instructions in issue descriptions, CI logs, and AGENTS.md files.

The plan is to see whether these injections can trick the agent into unauthorized actions (modifying protected files, accessing restricted info), then compare defenses such as prompt-based mitigations versus tool-level permission controls, measuring which reduce attacks without breaking legitimate tasks. The author is soliciting feedback and paper/resource suggestions, framing the goal as mapping where coding agents are weak rather than shipping yet another coding assistant.

Original post →

More from coding & agent

coding & agent channel →