tftsr-devops_investigation/docs/wiki/Security-Model.md
Shaun Arman 3588399dfd
Some checks failed
Test / rust-fmt-check (pull_request) Failing after 2m12s
Test / frontend-typecheck (pull_request) Successful in 2m23s
Test / frontend-tests (pull_request) Successful in 2m22s
Test / rust-clippy (pull_request) Successful in 3m55s
Test / rust-tests (pull_request) Successful in 5m10s
PR Review Automation / review (pull_request) Failing after 11m6s
feat(mcp): add MCP Server Support with TDD implementation
Adds full Model Context Protocol (MCP) server management, enabling the
AI assistant to discover and call tools from external MCP servers during
triage conversations.

Backend (Rust):
- rmcp 1.7.0 dependency (client + stdio + Streamable HTTP transports)
- Migration 018: mcp_servers, mcp_tools, mcp_resources tables with
  CHECK constraints for transport_type, auth_type, discovery_status
- src/mcp/ module: models, store, client, adapter, discovery, commands,
  transport/{stdio,http}
- AppState gains mcp_connections: Arc<TokioMutex<HashMap<...>>>
- .setup() hook auto-discovers enabled servers at startup
- 8 new Tauri commands wired into invoke_handler
- execute_mcp_tool_call: PII scan + mandatory audit_log before execution
- Auth values encrypted at rest via integrations::auth::encrypt_token();
  scrubbed before any frontend response

Frontend:
- MCPServers.tsx settings page (/settings/mcp) with server list,
  status badges, Discover Now, Add/Edit modal, enable/disable toggle
- tauriCommands.ts: McpServer, McpTool, McpServerStatus types + 8 cmds
- App.tsx: Plug icon, /settings/mcp route, sidebar nav entry

Tests (TDD): 15 new tests, all green
- 5 migration tests (written before migration, red → green)
- 5 store CRUD + encryption tests
- 5 adapter sanitization + conversion tests

Verification: 185/185 Rust, 94/94 Vitest, clippy -D warnings: 0
2026-05-23 16:23:48 -05:00

194 lines
6.9 KiB
Markdown

# Security Model
## Threat Model Summary
TFTSR handles sensitive IT incident data including log files that may contain credentials, PII, and internal infrastructure details. The security model addresses:
1. **Data at rest** — Database encryption
2. **Data in transit** — PII redaction before AI send, TLS for all outbound requests
3. **Secret storage** — API keys in Stronghold vault
4. **Audit trail** — Complete log of every external data transmission
5. **Least privilege** — Minimal Tauri capabilities
---
## Database Encryption (SQLCipher AES-256)
Production builds use SQLCipher:
- **Cipher:** AES-256-CBC
- **KDF:** PBKDF2-HMAC-SHA512, 256,000 iterations
- **HMAC:** HMAC-SHA512
- **Page size:** 16384 bytes
- **Key source:** `TFTSR_DB_KEY` environment variable
Debug builds use plain SQLite (no encryption) for developer convenience.
Release builds now fail startup if `TFTSR_DB_KEY` is missing or empty.
---
## Credential Encryption
Integration tokens are encrypted with AES-256-GCM before persistence:
- **Key source:** `TFTSR_ENCRYPTION_KEY` (required in release builds)
- **Key derivation:** SHA-256 hash of key material to a fixed 32-byte AES key
- **Nonce:** Cryptographically secure random nonce per encryption
Release builds fail secure operations if `TFTSR_ENCRYPTION_KEY` is unset or empty.
The Stronghold plugin remains enabled and now uses a per-installation salt derived from the app data directory path hash instead of a fixed static salt.
---
## PII Redaction
**Mandatory path:** No text can be sent to an AI provider without going through the PII detection and user-approval flow.
```
log file → detect_pii() → user approves spans → apply_redactions() → AI provider
```
- Original text **never leaves the machine**
- Only the redacted version is transmitted
- The SHA-256 hash of the redacted text is recorded in the audit log for integrity verification
- `pii_spans.original_value` is cleared after redaction to avoid retaining raw detected secrets in storage
- See [PII Detection](PII-Detection) for the full list of detected patterns
---
## Audit Log
Every external data transmission is recorded:
```rust
write_audit_event(
&conn,
action, // "ai_send", "publish_to_confluence", etc.
entity_type, // "issue", "document"
entity_id, // UUID of the related record
details, // JSON: provider, model, hashes, log_file_ids
)?;
```
The audit log is stored in the encrypted SQLite database. It cannot be deleted through the UI.
### Tamper Evidence
`audit_log` entries now include:
- `prev_hash` — hash of the previous audit entry
- `entry_hash` — SHA-256 hash of current entry payload + `prev_hash`
This creates a hash chain and makes post-hoc modification detectable.
**Audit entry fields:**
- `action` — what was done
- `entity_type` — type of record involved
- `entity_id` — UUID of that record
- `user_id` — always `"local"` (single-user app)
- `details` — JSON blob with hashes and metadata
- `timestamp` — UTC datetime
---
## Tauri Capabilities (Least Privilege)
Defined in `src-tauri/capabilities/default.json`:
| Plugin | Permissions granted |
|--------|-------------------|
| `dialog` | `allow-open`, `allow-save` |
| `fs` | `read-text`, `write-text`, `read`, `write`, `mkdir` — scoped to app dir and temp |
| `shell` | `allow-open` only |
| `http` | default — connect only to approved origins |
---
## Content Security Policy
```
default-src 'self';
style-src 'self' 'unsafe-inline';
img-src 'self' data: asset: https:;
connect-src 'self'
http://localhost:11434
https://api.openai.com
https://api.anthropic.com
https://api.mistral.ai
https://generativelanguage.googleapis.com;
```
HTTP is blocked by default. Only whitelisted HTTPS endpoints (and localhost for Ollama) are reachable.
---
## TLS
All outbound HTTP requests use `reqwest` with certificate verification enabled and a request timeout configured for provider calls.
CI/CD currently uses internal `http://` endpoints for self-hosted Gitea release automation on a trusted LAN. Recommended hardening: migrate runners and API calls to HTTPS with internal certificates.
---
## MCP Server Security
MCP server support introduces external tool execution capabilities. The following controls mitigate the associated risks.
### Auth Value Storage
- Auth tokens (API keys, bearer tokens, OAuth2 access tokens) are encrypted with **AES-256-GCM** before persistence in `mcp_servers.auth_value`.
- Encryption uses the same key derivation as integration credentials (`TFTSR_ENCRYPTION_KEY` → SHA-256 → 32-byte AES key).
- Random 96-bit nonce per encryption operation.
- Format: `base64(nonce || ciphertext || tag)`.
### Server-Side Response Scrubbing
- `list_mcp_servers` and all CRUD commands set `auth_value = None` before returning to the frontend.
- The encrypted ciphertext never reaches the WebView layer.
- Decryption only occurs internally when establishing a connection (discovery) or executing a tool call.
### Audit Trail
- `write_audit_event` is called **before** every MCP tool execution with:
- `action`: `"mcp_tool_call"`
- `entity_type`: `"mcp_tool"`
- `entity_id`: the tool key being invoked
- `details`: JSON containing server ID, tool name, and argument hash
- This provides a complete, tamper-evident record of all external tool invocations.
### PII Scan on Arguments
- Before dispatching a tool call, the arguments JSON is scanned through the PII detection pipeline.
- If PII is detected, a **non-blocking warning** is surfaced to the user.
- This prevents inadvertent leakage of credentials, email addresses, or IP addresses to external MCP servers.
### Stdio Transport Path Validation
- `build_stdio_transport()` rejects any `command` that is not an absolute path.
- This prevents:
- Path traversal attacks (e.g., `../../malicious`)
- Reliance on `$PATH` resolution which could be manipulated
- Unintended execution of relative-path binaries
### Network Boundaries
- HTTP transport uses `reqwest` with TLS certificate verification for HTTPS endpoints.
- stdio transport communicates only with locally spawned processes (no network exposure).
- MCP server URLs should be added to the Content Security Policy `connect-src` if fetched from the WebView layer.
### Cascade Deletes
- Removing an MCP server cascades to delete all associated `mcp_tools` and `mcp_resources` records.
- The live connection is also removed from the in-memory connection pool.
- No orphaned tool definitions can persist after server removal.
---
## Security Checklist for New Features
- [ ] Does it send data externally? → Add audit log entry
- [ ] Does it handle user-provided text? → Run PII detection first
- [ ] Does it store secrets? → Use Stronghold, not the SQLite DB
- [ ] Does it need filesystem access? → Scope the fs capability
- [ ] Does it need a new HTTP endpoint? → Add to CSP `connect-src`
- [ ] Does it add a new provider endpoint? → Avoid query-param secrets, use auth headers