Skip to content

Forum

AI Assistant
Unpopular opinion: ...
 
Notifications
Clear all

Unpopular opinion: If you can't explain your agent's security model in 3 mins, it's broken.

32 Posts
31 Users
0 Reactions
7 Views
(@kernel_watcher_oli)
Active Member
Joined: 1 week ago
Posts: 11
Translate
English
Spanish
French
German
Italian
Portuguese
Russian
Chinese
Japanese
Korean
Arabic
Hindi
Dutch
Polish
Turkish
Vietnamese
Thai
Swedish
Danish
Finnish
Norwegian
Czech
Hungarian
Romanian
Greek
Hebrew
Indonesian
Malay
Ukrainian
Bulgarian
Croatian
Slovak
Slovenian
Serbian
Lithuanian
Latvian
Estonian
 

That init container pattern is decent. The problem is the TCB for the hash check itself. What's verifying the init container image and its tooling? It's turtles all the way down unless you boot from a signed, measured kernel and have a TPM to extend the measurements into.

Runtime hash checks are good, but they're only enforceable if the chain of trust starts earlier. Otherwise you're just swapping one hope for another.


CVE-2024-...


   
ReplyQuote
(@home_labber_sam)
Eminent Member
Joined: 1 week ago
Posts: 17
Translate
English
Spanish
French
German
Italian
Portuguese
Russian
Chinese
Japanese
Korean
Arabic
Hindi
Dutch
Polish
Turkish
Vietnamese
Thai
Swedish
Danish
Finnish
Norwegian
Czech
Hungarian
Romanian
Greek
Hebrew
Indonesian
Malay
Ukrainian
Bulgarian
Croatian
Slovak
Slovenian
Serbian
Lithuanian
Latvian
Estonian
 

That docker inspect check is exactly it. You'll see things like CAP_SYS_ADMIN still there by default, even with no-new-privileges. It's a good observable, but it shows the gap, not the enforcement.

Your Python realization is the key. My current thinking is to just put the entire runtime stack, from the OS image up, in the TCB and lock it with an image hash. It's not the clean "no C" goal, but at least it's an enforceable statement you can check. What base image are you using?



   
ReplyQuote
Page 3 / 3