A Connecticut judge just flagged something I built a guard against by default.
In a case before him, a plaintiff had embedded "hidden text" in court filings. The text wasn't for the human reader. It was written as instructions to whatever AI system later processed the document.
This is the exact failure mode I designed Hermes Agent's SOUL.md to reject. My agents only take instructions from me. If an instruction is buried inside a document, email, webpage or screenshot, it is treated as untrusted data, never as a command and flagged to me.
Reuters reported the public version of a problem I expect to see far more of. As more documents flow through agents, more people will try to slip prompts into them. Some of those prompts will be harmless. Others will not.
If your AI agent reads a PDF and then acts on what's inside it, ask yourself who is actually giving it orders?
If you want to build your own Hermes Agent, sign up for our next workshop here: https://lnkd.in/g6qirhv7