Prismor is agent agnostic and works with Claude Code, Codex, Langchain, Vercel AI SDK and 30+ more agent frameworks
Prismor judge API is free until the end of the month
let your agent setup prismor in 30 seconds from this repo
PrismorSec/prismor
6 likes · 61 views
Companies will keep blocking AI tools and employees will keep using it
I work in a pharma company and we're not allowed to use agents. I think every employee knows how useful agents are but companies will keep fearing security, as emploiyees we are all managers of our agents
linkedin is full of AI slop but I found this post cool where 3 fundamental questions all companies should strive to answer:
1. What agents are running?
2. Which devices and employees permission these agents?
3. Can we verify and control agent actions?
and they shared this open source which I'm absolutely loving while using it this weekend.
It was able to give me ful control over my agents fetching untrusted content, sharing my emails and managing secrets.
Is anyone else feeling restricted with their company policy around using AI agents?
49 points · 23 comments
This tweet is a scary truth
Credit due to Steven\^
I'll be honest I faced a similar issue with claude when building out an oss project related to ai safety. I realized that giving claude code unrestircted network access and creds is a massive privacy risk
what if we can have an automatic way of redacting any secret sent to AI
for example your AWS creds might be ABC but secret redacted can be DEF to the AI agent like claude and then replaced again with ABC for connection with the server?
and what if we can connect to external hosts only based on the permission we gave to the agents, like a fly on the wall which makes agent actions secure
There are 28 researchers who spent last 10 months working on it with me and we would be grateful to the community for any feedback on our open source project prismor
658 points · 130 comments
This tweet is a scary truth
Credit due to Steven\^
I'll be honest I faced a similar issue with claude when building out an oss project related to ai safety. I realized that giving claude code unrestircted network access and creds is a massive privacy risk
what if we can have an automatic way of redacting any secret sent to AI
for example your AWS creds might be ABC but secret redacted can be DEF to the AI agent like claude and then replaced again with ABC for connection with the server
and what if we can connect to external hosts only based on the permission we gave to the agents, that's something we've been researching and working with the community of 28 contributors so we can build a fly on the wall which makes agent actions secure
I have spent my last 10 months working on it and would be grateful to the community for any feedback on our open source project prismor
199 points · 22 comments
@vishalsingh2972 terrifying is fair. the agent already holds the keys, so a scan of its dependencies doesn't cover it. what matters is the moment it reads .env or runs a command mid-run. that's where you want a guard. do checkout :PrismorSec/prismor
3 likes · 20 views
who did this?
PrismorSec/prismor t.co
3 likes · 1 repost · 139 views
Based on user feedback. New this week on Prismor
PrismorSec/prismor
4 likes · 1 repost · 141 views
ALMOST THERE TO STARS NOT FOUND
PrismorSec/prismor t.co
10 likes · 421 views
Mentions
Prismor is agent agnostic and works with Claude Code, Codex, Langchain, Vercel AI SDK and 30+ more agent frameworks Prismor judge API is free until the end of the month let your agent setup prismor in 30 seconds from this repo PrismorSec/prismor
Companies will keep blocking AI tools and employees will keep using it I work in a pharma company and we're not allowed to use agents. I think every employee knows how useful agents are but companies will keep fearing security, as emploiyees we are all managers of our agents linkedin is full of AI slop but I found this post cool where 3 fundamental questions all companies should strive to answer: 1. What agents are running? 2. Which devices and employees permission these agents? 3. Can we verify and control agent actions? and they shared this open source which I'm absolutely loving while using it this weekend. It was able to give me ful control over my agents fetching untrusted content, sharing my emails and managing secrets. Is anyone else feeling restricted with their company policy around using AI agents?
This tweet is a scary truth Credit due to Steven\^ I'll be honest I faced a similar issue with claude when building out an oss project related to ai safety. I realized that giving claude code unrestircted network access and creds is a massive privacy risk what if we can have an automatic way of redacting any secret sent to AI for example your AWS creds might be ABC but secret redacted can be DEF to the AI agent like claude and then replaced again with ABC for connection with the server? and what if we can connect to external hosts only based on the permission we gave to the agents, like a fly on the wall which makes agent actions secure There are 28 researchers who spent last 10 months working on it with me and we would be grateful to the community for any feedback on our open source project prismor
This tweet is a scary truth Credit due to Steven\^ I'll be honest I faced a similar issue with claude when building out an oss project related to ai safety. I realized that giving claude code unrestircted network access and creds is a massive privacy risk what if we can have an automatic way of redacting any secret sent to AI for example your AWS creds might be ABC but secret redacted can be DEF to the AI agent like claude and then replaced again with ABC for connection with the server and what if we can connect to external hosts only based on the permission we gave to the agents, that's something we've been researching and working with the community of 28 contributors so we can build a fly on the wall which makes agent actions secure I have spent my last 10 months working on it and would be grateful to the community for any feedback on our open source project prismor
@vishalsingh2972 terrifying is fair. the agent already holds the keys, so a scan of its dependencies doesn't cover it. what matters is the moment it reads .env or runs a command mid-run. that's where you want a guard. do checkout :PrismorSec/prismor
who did this? PrismorSec/prismor t.co
Based on user feedback. New this week on Prismor PrismorSec/prismor
ALMOST THERE TO STARS NOT FOUND PrismorSec/prismor t.co