This tweet is a scary truth
retained text
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](https://github.com/PrismorSec/prismor)
document fetched 2026-10-27 00:41 · canonical https://reddit.com/r/ClaudeCode/comments/1xg1zev/this_tweet_is_a_scary_truth