Who Is Prompt-Injecting Coding Agents? A Scan of 86 Million GitHub Items

Gabi Nakibly

Video

Abstract:

A hidden HTML comment in a GitHub issue can hijack a coding agent. This has been repeatedly shown in demos and research PoCs. But how often are attackers in the wild actually trying it, and what do they want? We launched the first GitHub-wide scan to hunt for indirect prompt injections across 86 million GitHub items: issues, PRs, comments and READMEs. We extracted hidden text (e.g., HTML comments) then screened it with novel heuristics and a two-stage LLM classifier. We then manually reviewed every flagged candidate. We identified prompt injections in roughly 2.5 of every 100K GitHub items. A key finding is that injections serve diverse objectives: orchestration of project agents, constrain visiting bots, growth hacking, and malicious exfil data. We’ll dissect real injections, map their objectives, and explain our detection blind spots. We release the detector, prompts, and taxonomy so others can investigate who is steering coding agents and why.

Bio:

Dr. Gabi Nakibly is a Distinguished Researcher at Radware, where his responsibilities include leading Agentic AI Protection R&D, and a Senior Lecturer in the Technion’s Department of Computer Science. He was twice a Visiting Scholar at Stanford University’s Computer Security Lab and previously served as CTO of a national defense research laboratory. His work spans AI, application, and network security. He has presented six times at Black Hat and at conferences including RSA, NDSS, and USENIX Security.

Time and Place

Monday, October 26, 4:30pm
CoDa E201 & Zoom