jailbreaks are not magic, they are mostly just persistence and context
spent a week red teaming our assistant. the successful attacks were not clever, they were just patient and used the model own context against it. defense in depth or bust.
this should be pinned
this is gonna be obsolete in a month but useful now
did you compare against the obvious baseline?
i would add: log everything, you will thank yourself later
good post but the title oversells it a little
wait how did you actually get that working
ok this finally made it click for me, thank you
the eval first mindset is underrated, nice to see it here
saving this, exactly what I needed today
honestly wild that this works at all
ok but does it survive a hostile user
the real lesson here is to not trust the happy path
skeptical but bookmarking to test friday
do you have a repo or gist? would love to poke at it
thanks, this saved me probably a full day
any gotchas you ran into setting it up?
tried it for an afternoon, not sold yet
the prompt injection example is terrifyingly simple
how are you handling auth for the tool calls?
tried it for an afternoon, not sold yet
this is gonna be obsolete in a month but useful now
this is a really clean mental model, thanks
underrated post, more people should see this
tried it for an afternoon, not sold yet
lol the agent did this to me too and I almost shipped it
you just described my entire last sprint
the eval first mindset is underrated, nice to see it here
agree with the conclusion, not the reasoning