2013488 - Security layer - Implement tainting for all actions/input and enforce policy based on those two criteria

Status: RESOLVED WONTFIX

(Core :: Machine Learning: General, task)

Description

[Frederik Braun [:freddy]]

1. Implement Global Taint State per context
2. Implement Input Taint Injection & Tagging.
3. Define Taint Propagation Rules for LLM output.

Once that is done:

• Implement Action Sensitivity Mapping (Sinks).
• Implement Policy Enforcement Logic
• Implement/Experiment with different (user visible) enforcement results (url filtering, compacting context, asking for permission)

Comment 1

[Greg Tatum [:gregtatum]]

I'm taking this out of blocking the MVP as it's a bit of a duplicate of Bug 2020825 (which I filed). However, there are more details here that aren't in the security alignment doc. There are a bunch of dependent bugs so I'm hesitant to do any more work moving things around until after the MVP.

Comment 2

[Greg Tatum [:gregtatum]]

This isn't in the security alignment doc, and pieces of this work have been completed. We can file new bugs for more well defined work here. We implemented SecurityProperties and other mechanisms for tracking "tainting flags" and their enforcement. If we want a more elegant design, we can specify that work.