CLAIM-2026-10-09-06
The paper describes planting malicious prompts in benign community rule files such as AGENTS.md or .cursorrules so that coding agents swap legitimate dependencies for attacker-controlled packages, and reports transferability across models and agent frameworks.
Why it matters
Carries the 2026-10-09-the-hook-that-let-it-through edition's argument about a control plane that defaults to permissive.
Sources
- arxiv2026package