ctippett

This sounds intriguing. Can you expand on what your thinking is or link to somewhere I could read more?

They wouldn't. A device like the Pineapple broadcasts its own copycat wireless signal and tricks unsuspecting clients to connect to it thinking it's the real one.

Edit: Holy moly we all jumped on this comment at the same time didn't we.

This was(/is?) a great write-up. Would the vulnerability enabled through Taildrop still work if the compromised machine was tagged or would the attacker's control of the... control plane make this additional safeguard redundant?

Tags are a key part of managing ACLs. If you're concerned about an untrustworthy node on your tailnet, use tags. Even if you forget to add a tag to a newly added node (API keys can be set to automatically assume a tag when used fwiw), if your other devices are tagged then they're ineligible for taildrop.

Taildrop only works with devices that are yours, i.e. you can't send or receive anything from a device that's either tagged or owned by someone else. The only way malware is getting onto your NAS (via Taildrop) is if you're the one sending it there.

Sometimes when I come up with analogies they miss the mark. Yours was on point.

I like the idea, but its a surefire way of bloating the context. Although I guess you'd want a subagent to do the search.

  > In theory, theory and practice are the same. In practice, they're not.

This is such beautiful rhetoric.

Unfortunately not, the preexec event happens asynchronously from the actual execution of the command. You can't manipulate or otherwise log the output of what's run, all you get is a copy of the command text.

You could probably achieve something like that using Fish shell and their fish_preexec and fish_postexec event hooks.

Edit: Nope, looks like you can't. You only get the invoked command and no output.

This. It's also while you'll sometimes get a mangled Getty Images watermark on some image generations, or a logo in the bottom left corner. If it's a prominent feature in the training dataset it'll show up, exactly how these models are supposed to work.

The 'bug' here is whatever post-processing step or system prompt is in place to steer the model away from doing this.

Perhaps not an HN username, but several Tippett's seem to think they all share my gmail address based on the invitations, council minutes, school notices and bank statements that get sent my way.

The best was another famous Tippett of NFL fame sending me a $500 gift voucher for their son's birthday.