Inspired by the recent Openai’s Huggingface incident, I have started creating a new very small programming language (Leashterm) especially for agents. So, this language written in Rust, is not meant for human programmers, but rather for agents that must be contained to not break rules, and such that the language itself constrains the capabilities of the agents.
In this way, security is build in, in the language. We performed some tests, comparing chatgpt written Python codes versus Leashterm. We show in cases that the Python scripts break rules to obtaim their goals, and also Leashterm scripts, but the Leashterm scripts are prevented from executing these not permitted actions.
Inspired by the recent Openai’s Huggingface incident, I have started creating a new very small programming language (Leashterm) especially for agents. So, this language written in Rust, is not meant for human programmers, but rather for agents that must be contained to not break rules, and such that the language itself constrains the capabilities of the agents. In this way, security is build in, in the language. We performed some tests, comparing chatgpt written Python codes versus Leashterm. We show in cases that the Python scripts break rules to obtaim their goals, and also Leashterm scripts, but the Leashterm scripts are prevented from executing these not permitted actions.