Security fix
With the ReAct runner (runner: react), ReactConverter fell back to Python eval() on the model's Action: text whenever it was not valid JSON. A prompt-injected action (from a web page, tool result or document the agent reads) could execute arbitrary Python on the host. Actions are now parsed with ast.literal_eval and must be a dict with name and arguments. Reported through TSRC.
The default openai runner is not affected. Upgrade if you use the ReAct runner.