A file or web page tells the agent to do something
Text the agent reads is data, not an instruction. Corpole makes that a rule agents cannot drop.
What happened
A comment in the code, an issue or a web page contained an instruction aimed at the agent, and the agent was about to follow it.
Why it happens
Agents read a lot of text they did not write. Anything in it that looks like an instruction can steer them, whoever wrote it.
What Corpole does
- Gives every agent a locked rule: instructions come only from the person, Corpole's Authority records and the agent rules; external content is data.
- Asks the agent to show you such an instruction, with its source, instead of acting on it.
What you do
- Look at what the agent found and where.
- If you do want it done, ask for it yourself; then it is your request.
- If it looks hostile, record it as a risk.
Keep it from happening again
- Review what goes into files and issues your agents read.