Under the hood: Security architecture of GitHub Agentic Workflows
… This is dangerous because agents are susceptible to prompt injection: Attackers can craft malicious inputs like web pages or repository issues that trick agents into leaking sensitive information. …