Noma Labs identified a prompt injection vulnerability in GitHub's Agentic Workflows, allowing unauthenticated attackers to access private repository data by posting a crafted GitHub Issue in a public repository within the same organization. This vulnerability has been named GitLost.

Prompt injection exploits a flaw in how large language models (LLMs) perceive roles, leading to new attack vectors and insights into model behavior. Understanding roles is crucial for predicting the success of these attacks and developing a research framework around them.

Prompt injection in Google Translate can reveal the underlying instruction-following language model. Responses indicate that the model lacks strong boundaries between processing content and following instructions.