🟢 #155 fix: validate request body in POST /projects/:projectId/chat +118 -8 19d ago by
Shawnaldinho ↗ analysis ↗ GitHub ## Summary
Shawnaldinho/mike
A security-minded fork of Mike, hardening the chat backend against prompt injection, unchecked inputs, and unmetered usage.
This is Shawnaldinho's hardening pass over Mike. It doesn't rebrand the product or add a new surface for users - the work sits under the hood, aimed at making the assistant safer and more accountable when it handles real requests.
The through-line is defensive engineering. The tracked threads close prompt-injection holes where hostile text could sneak into the assistant's instructions, tighten what the chat endpoint is willing to accept from a client, add logging that lets an error be traced back to the request that caused it, and turn a decorative "credits remaining" figure into a budget that actually counts and enforces usage.
Who's behind it? Only the handle: Shawnaldinho. There's no rebrand, no stated niche, and no deployment story to report - this reads as an engineer shoring up the foundations rather than pitching a new flavor of Mike. If the security angle interests you, the details are worth a click through to GitHub.
What's in it
- Prompt-injection defenses Two threads aim at the same risk: keeping user-supplied text - filenames, folder paths, workflow titles, tool results - from being read as instructions and hijacking the assistant.
- Stricter request validation The project chat endpoint stops taking the request body on faith and checks that what a client sends actually has the expected shape before acting on it.
- Structured logging Per-request identifiers and one access line per response, so an error a user sees can be correlated back to the request that produced it.
- Enforced usage budgets The "credits remaining" number becomes real - usage gets counted after each call and checked before the next, instead of sitting at zero and enforcing nothing.
Direction
securityinfrastructure
Activity
🟢 #156 feat: structured JSON logger with per-request id and access log +149 -5 19d ago by
Shawnaldinho ↗ analysis ↗ GitHub ## Summary
🟢 #158 security: spotlight untrusted content + ship an honest threat model +612 -17 19d ago by
Shawnaldinho ↗ analysis ↗ GitHub ## Background
⛔ #157 fix: enforce monthly message-credit limit before chat LLM calls +125 -3 21d ago by
Shawnaldinho ↗ analysis ↗ GitHub ## Summary
⛔ #154 fix: sanitize untrusted strings interpolated into LLM prompts +48 -9 2mo ago by
Shawnaldinho ↗ analysis ↗ GitHub ## Summary
📝 Shawnaldinho spots a filename that could hijack the assistant, then lets the fix lapse 0 commits — securitychat-ui draft
A document named the right way could smuggle instructions past the AI and make it ignore its own rules.
The endpoint behind project chat used to trust whatever a client sent it. Shawnaldinho proposes checking the mail before it opens the box.
📝 Shawnaldinho wants every broken chat to come with a tracking number 0 commits — infrastructure draft
A proposal to give the backend a proper paper trail so a failure a user sees can be traced back to the exact line that caused it.
📝 Shawnaldinho tries to make Mike's credit meter actually count 0 commits — infrastructureanalytics draft
The "credits remaining" figure every Mike user sees was decorative - it never dropped and never stopped anyone.
📝 Shawnaldinho builds a prompt-injection defense that admits where it stops 0 commits — security draft
A proposed shield against documents that try to hijack the AI - paired with an unusually honest note about what it can't do.
Threads of work (detailed view)
Shawnaldinho builds a prompt-injection defense that admits where it stops
A proposed shield against documents that try to hijack the AI - paired with an unusually honest note about what it can't do.
Shawnaldinho tries to make Mike's credit meter actually count
The "credits remaining" figure every Mike user sees was decorative - it never dropped and never stopped anyone.
Shawnaldinho wants every broken chat to come with a tracking number
A proposal to give the backend a proper paper trail so a failure a user sees can be traced back to the exact line that caused it.
Shawnaldinho bolts a lock onto the project-chat door
The endpoint behind project chat used to trust whatever a client sent it. Shawnaldinho proposes checking the mail before it opens the box.
Shawnaldinho spots a filename that could hijack the assistant, then lets the fix lapse
A document named the right way could smuggle instructions past the AI and make it ignore its own rules.
Pull requests (detailed view)
🟢 Open (3)
by
Shawnaldinho · opened 2mo ago by
Shawnaldinho · opened 2mo ago by
Shawnaldinho · opened 2mo ago ⛔ Closed without merge (2)
by
Shawnaldinho · opened 2mo ago · closed 21d ago by
Shawnaldinho · opened 2mo ago · closed 2mo ago