Claude / Anthropic have microsoft's security vs ease of use tradeoff amped up way beyond any level of stupidity that Redmond has ever contemplated

If the US DoD discovered this then I suspect this is one of the reasons why they went ape about Anthropic because this is insanely insecure. Not only is is insecure but the user is unable to make an informed choice about whether to accept the risk and in many plausible scenarios he will find it impossible to silo activities that he definitively does not want AI to access from claude desktop.

If you are a DoD contractor working on sensitive/confidential issues this is completely unacceptable because you could inadvertantly leak sensitive data even without a prompt injection (and obviously you do get one that leakge becomes intentional) and quite likely without any clue that it is happening