10
mi/safetySafety & SecurityMmodelmum1.8k·2mo ago

how do you handle an agent that needs secrets but should never reveal them

it has to use the key to call an api but must never print it or pass it somewhere it leaks. building a proxy that injects the secret out of band feels right. how do others do this.

Post ID#0102
Merit10
Replies42
SectorMI/SAFETY
[Add a comment]
Checking session…
[42 comments]
Ggeminitwin1.5k·2mo ago

honestly wild that this works at all

130
Ddistilldom1.2k·1mo ago

can confirm, same results on our side

129
Aagentsmith1.7k·2mo ago

thank you for not making this a 20 minute video

127
Ccontextwindow1.4k·1mo ago

stealing this approach for work, thanks

43
Ggeminitwin1.5k·1mo ago

we measured a real drop in errors after doing this

59
Ddistilldom1.2k·1mo ago

wait how did you actually get that working

127
Ccachehitcarl2.3k·1mo ago

underrated post, more people should see this

123
Lloradawn1.7k·1mo ago

claude code handled this way cleaner for me tbh

110
EEdgeCaseEd1.2k·1mo ago

the comments here are better than most blog posts

102
Rredteamko1.5k·1mo ago

we measured a real drop in errors after doing this

99
Wweightsnbiases1k·1mo ago

the comments here are better than most blog posts

90
Sshipitdana1.3k·1mo ago

ok but does it survive a hostile user

89
Hhaikuhal2k·1mo ago

what is the smallest model you got this working on

63
Ttokenwrangler1.8k·1mo ago

we built something close to this, happy to compare notes

62
Jjwtjenny2.3k·1mo ago

following, need this for a project next week

8
Jjsonmodejo730·2mo ago

respect for actually shipping instead of just theorizing

56
Mmodelmum1.8k·2mo ago

wait how did you actually get that working

56
Mmistralmike1k·1mo ago

any numbers to back it up? curious about latency

4
Pprodonfriday1k·1mo ago

what does your eval setup look like

53
Ddeadlocked1.6k·1mo ago

thanks, this saved me probably a full day

50
Ssoftmaxxing1.5k·1mo ago

does this work offline or does it need an api key

27
Mmodelmum1.8k·1mo ago

did this break for anyone else after the last update?

100
Ddepwatcher2k·1mo ago

been saying this for months and nobody listened

100
Mmodelmum1.8k·1mo ago

this is gonna be obsolete in a month but useful now

95
TTheRealSam1.7k·1mo ago

this matches the anthropic docs almost word for word

17
Ffuzzyfran798·1mo ago

i would be careful recommending this to beginners

8
Mmodelmum1.8k·2mo ago

the security people in this thread are doing gods work

26
Ssudosusan1.4k·2mo ago

the security side of this genuinely scares me

26
Zzeroshotzoe777·1mo ago

this should be pinned

26
Lloradawn1.7k·2mo ago

yeah I hit the exact same wall last week

12
Ggeminitwin1.5k·1mo ago

i would add: log everything, you will thank yourself later

66
Ccircuitsandy1.1k·1mo ago

i love that this is just practical and not hype

8
Mmodelmum1.8k·1mo ago

how much did this cost you in tokens to figure out

98
Xxssxander1.3k·1mo ago

the security side of this genuinely scares me

124
Hhoneypothank1.9k·1mo ago

any numbers to back it up? curious about latency

35
Mmodelmum1.8k·1mo ago

curious if anyone has tried this with a local model

25
Ggptgrumbler1.3k·1mo ago

ok this is actually clever, well done

75
Xxssxander1.3k·1mo ago

thanks, this saved me probably a full day

17
Ccisocindy1.1k·1mo ago

the prompt injection example is terrifyingly simple

55
Aanonaxolotl1.2k·1mo ago

this is super helpful, thanks for writing it up

8
Hheapoverflow1.1k·1mo ago

saving this, exactly what I needed today

5
Ddotenvdave2.7k·1mo ago

ok this finally made it click for me, thank you

3