⛔ #29 fix(security): remove persisted Claude raw stream log +4 -9 6mo ago by
fayerman-source ↗ analysis ↗ GitHub ## Summary
fayerman-source/mike
A security-minded Mike fork focused on safer legal AI evaluation and stronger project-level privacy boundaries.
fayerman-source is shaping Mike for legal teams that need to test AI-assisted document work without treating sensitive material casually. The fork focuses on keeping conversational and document context out of default plaintext retention, while preserving an opt-in route for troubleshooting.
It also tightens the separation between projects when people reorganise folders or documents, and adds practical guidance for firms trying Mike locally with disposable infrastructure, synthetic material, and limited-spend credentials. The direction is cautious, self-hosting-friendly legal AI rather than a visible product redesign.
What's in it
- Private-by-default AI activity Raw AI streaming material is no longer written to a plaintext log by default; diagnostics require an explicit opt-in.
- Project boundary protection Folder and document moves are checked so one project cannot reach into another project's workspace.
- Safer evaluation guidance A dedicated local-testing path helps legal evaluators use disposable setups, synthetic documents, capped keys, and safer credential handling.
Direction
securityworkflow
Activity
## Summary
✅ #28 fix(projects): validate folder ownership before folder mutations +28 -3 6mo ago by
fayerman-source ↗ analysis ↗ GitHub ## Summary
📝 fayerman-source stops Mike from quietly writing privileged material to disk 1 commit 6mo ago securitycompliance draft
Every AI exchange was being saved to a plaintext file on the server, forever. Now it isn't.
📝 fayerman-source walls off one project's folders from another 1 commit 6mo ago securitymulti-tenant draft
A handful of folder and document actions trusted IDs from the browser without checking they belonged to the project you were actually in - and now they don't.
fayerman-source added a dedicated guide for evaluating Mike with sensitive legal workflows, linked from the setup README so it sits on the path a first-time self-hoster actually follows. The guide pushes toward disposabl…
Threads of work (detailed view)
fayerman-source stops Mike from quietly writing privileged material to disk
Every AI exchange was being saved to a plaintext file on the server, forever. Now it isn't.
Raw Claude stream logging gated behind DEBUG_LLM_STREAM flag
fayerman-source proposed removing the unconditional write of `claude-raw-stream.log` and replacing it with a debug flag. The concern was concrete: raw stream events carry user prompts and document context, and writing every event to an unbounded plaintext file is unnecessary data retention in a legal workflow context.
fayerman-source walls off one project's folders from another
A handful of folder and document actions trusted IDs from the browser without checking they belonged to the project you were actually in - and now they don't.
fayerman-source adds project boundary checks on folder and document mutations
fayerman-source found three places in the folder/document mutation APIs where a client-supplied folder ID was trusted without verifying it belonged to the calling project. In a multi-tenant legal document tool that's an IDOR-shaped gap.
Safe local testing guide added to onboarding path
fayerman-source added a dedicated guide for evaluating Mike with sensitive legal workflows, linked from the setup README so it sits on the path a first-time self-hoster actually follows. The guide pushes toward disposable infrastructure, synthetic documents, spend-capped API keys, and keeping service-role secrets off the frontend.
Pull requests (detailed view)
✅ Merged (2)
by
fayerman-source · opened 6mo ago · merged 6mo ago by willchen96 by
fayerman-source · opened 6mo ago · merged 6mo ago by willchen96 ⛔ Closed without merge (1)
by
fayerman-source · opened 6mo ago · closed 6mo ago