Show HN: Taster – Plain-English rules your agent can't forget or ignore (github.com)

Taster, a new Python library, enhances AI agent safety by ensuring all tool calls are checked against user-defined rules written in plain English before execution. This innovative approach addresses significant vulnerabilities that can arise when AI agents inadvertently execute harmful commands, like deleting sensitive information or leaking private data. By blocking actions that conflict with user-set principles, such as "never delete more than one item at a time" or "never touch production," Taster minimizes the risk of untrustworthy or malicious behavior from AI agents. This solution is notable because it positions user-defined guidelines outside the agent, making them nearly impossible to forget or dispute, whereas traditional rule engines may overlook subtle variations in harmful commands. Taster operates in a "shadow mode" initially, allowing users to monitor potential blocks without enforcing them immediately, thus facilitating seamless integration into existing AI systems. Its design also considers performance, handling checks quickly to prevent disruptions in agent operations. With Taster, the AI/ML community gains a critical tool for establishing trust and reliability in autonomous agents, pushing towards safer AI implementations.