This doesn’t solve hostile agents. This solves hostile or compromised inference providers. You really don’t want your secrets in the logs of a random AI provider through OpenRouter or even in Anthropic logs.
Would love to explore this train of thought and what we can do about it.