prompt injection defenses are not solved, stop selling them as solved
every vendor claims their filter stops injection. I have broken several in an afternoon. treat detection as one layer, never the whole defense. assume it gets through.
the eval first mindset is underrated, nice to see it here
what is the smallest model you got this working on
the real lesson here is to not trust the happy path
ok now do the version that handles errors
what is the failure mode when the tool call times out?
appreciate you sharing the failures too, not just the wins
tried, failed, tried again, finally works, can confirm
what model were you running for this?
this matches the anthropic docs almost word for word
the security side of this genuinely scares me
the eval first mindset is underrated, nice to see it here
ok now do the version that handles errors
the diagram alone is worth the read
more posts like this please
claude code handled this way cleaner for me tbh
genuinely useful, rare these days
this is the kind of post I come here for
honestly wild that this works at all
we got burned by this same thing in january