Invisible Ink for Robots: Hidden Unicode Tags Turn an OpenAI Security Skill Into a Backdoor

What happened

A security researcher showed that hidden Unicode Tag characters could be inserted into a legitimate security-best-practices Skill from OpenAI's curated Skills project, so an agent carried out instructions that were invisible to a human reviewer. In the Claude Code demo, auto-approval was on and the agent ran a shell command. The researcher later reported that Claude Code refused invisible-Unicode content on a February 10 retry, and no real-world harm is described.