NEW VULNERABILITY IN GITHUB COPILOT AND CURSOR: HOW HACKERS CAN WEAPONIZE CODE AGENTS

Summary from the TLDR InfoSec newsletter

Researchers from Pillar Security have discovered a new prompt injection technique on Cursor and GitHub Copilot. These systems use rules files to guide the prompting for the LLM, which is often distributed by open-source projects or from other untrusted locations. This attack involves utilizing invisible Unicode characters to hide malicious instructions for the LLM that can be used to instruct it to poison the generated code.