The risk isn't from the AI labs. It's from malicious attackers who sneak instructions to coding agents that cause them to steal your data, including your environment variable secrets - or cause them to perform destructive or otherwise harmful actions using the permissions that you've granted to them.