# Security Review — AI Gateway
**Date:** 2026-04-02
**Reviewer:** Claude Code (automated static analysis)
**Scope:** Full codebase (`d:\Antigravity Apps\ai_gateway`)
---
## Summary
| Severity | Count |
|---|---|
| Critical | 4 |
| High | 13 |
| Medium | 10 |
| Low | 9 |
| Informational | 3 |
| **Total** | **39** |
---
## Immediate Action Items
> Do these before any other work.
1. **Separate `API_KEY` and `SECRET_KEY`** — they are currently set to the same value; a leaked API key allows forging admin JWTs.
2. **Remove hardcoded fallback credentials** in `app/core/config.py` (`"admin123"`, `"storyline-secret-key-123"`).
3. **Add rate limiting to login** (`/auth/login`) — currently bruteforceable.
4. **Consider runtime secret injection** (Docker secrets, AWS Secrets Manager) so credentials are never written to disk at all.
---
## Critical
### C-1 — Plaintext Secrets in `.env` File
**File:** `.env` lines 1–10
Multiple live credentials stored in plaintext. The `.env` file is correctly listed in `.gitignore` and is **not committed to the repository**, so there is no git history exposure. However, the secrets still warrant attention:
| Variable | Issue |
|---|---|
| `API_KEY` | Same value as `SECRET_KEY` — two independent secrets should not share a value |
| `GOOGLE_API_KEY` | Live Google API key — should be scoped to minimum required APIs/quotas |
| `DATABASE_URL` | Full Supabase connection string with embedded password |
| `ADMIN_PASSWORD` | Plaintext in a file on disk; no rotation mechanism |
| `SECRET_KEY` | Identical to `API_KEY` — compromising one compromises both |
**Remaining concerns:**
- Anyone with filesystem access to the server (compromised container, misconfigured volume mount, log leak of `DATABASE_URL`) gets all credentials at once.
- `SECRET_KEY` and `API_KEY` being the same value means a leaked API key also forges JWTs.
**Fix:** Use a secrets manager or environment injection at runtime (Docker secrets, AWS Secrets Manager, Doppler) so credentials are never written to disk. At minimum, use separate values for `API_KEY` and `SECRET_KEY`.
---
### C-2 — Weak Default Credentials in Config
**File:** [app/core/config.py](app/core/config.py) lines 8, 13, 16
```python
API_KEY: str = os.getenv("API_KEY", "storyline-secret-key-123")
SECRET_KEY: str = os.getenv("SECRET_KEY", "your-super-secret-key-change-me")
ADMIN_PASSWORD: str = os.getenv("ADMIN_PASSWORD", "admin123")
```
If environment variables are absent, the app falls back to trivially guessable values.
**Fix:** Remove all default values. Raise a `RuntimeError` on startup if these are not set.
---
### C-3 — Fully Permissive CORS
**File:** [app/main.py](app/main.py) lines 49–56
```python
CORSMiddleware(
allow_origins=["*"],
allow_credentials=True,
allow_methods=["*"],
allow_headers=["*"],
)
```
`allow_origins=["*"]` combined with `allow_credentials=True` is both a browser-rejected combination and a clear signal of over-permissive intent. Any domain can make cross-origin requests.
**Fix:**
```python
allow_origins = os.getenv("CORS_ORIGINS", "https://your-frontend.com").split(",")
```
---
### C-4 — Hardcoded Admin Credentials with No Password Hashing
**File:** [app/api/endpoints/auth.py](app/api/endpoints/auth.py) lines 14–26
```python
if form_data.username == "admin" and form_data.password == settings.ADMIN_PASSWORD:
```
- Password is plaintext comparison — `passlib` is a dependency but never used.
- No rate limiting on the login endpoint → bruteforceable.
- Single static admin account with no database backing.
**Fix:** Use a proper user table with bcrypt-hashed passwords. Add `@limiter.limit("5/minute")` to the login route.
---
## High
### H-1 — Debug Print Statements Leak Auth Info in Production
**File:** [app/api/deps.py](app/api/deps.py) lines 23–28
```python
print(f"DEBUG: API Key received (prefix): {api_key[:5]}...")
```
First 5 chars of each API key are printed to stdout on every authenticated request. `ENVIRONMENT=production` is set in `.env`.
**Fix:** Remove all `print()` debug statements. Use structured logging gated on `DEBUG` level.
---
### H-2 — Request Logging Leaks URLs and Origins
**File:** [app/main.py](app/main.py) lines 40–47
Every incoming request, including URL (which may contain query-param secrets) and `Origin` header, is logged at INFO level in production.
**Fix:** Gate debug logging behind `settings.ENVIRONMENT == "development"`. Sanitize URLs before logging.
---
### H-3 — No Rate Limiting on Admin Endpoints
**File:** [app/api/endpoints/admin.py](app/api/endpoints/admin.py) lines 24–120
`POST /modules`, `GET /modules`, `POST /modules/{id}/rotate`, `DELETE /modules/{id}` — none have `@limiter.limit()`. Regular API endpoints use `settings.RATE_LIMIT` but admin does not.
**Fix:** Add `@limiter.limit("10/minute")` or stricter to all admin routes.
---
### H-4 — Duplicate Admin Endpoints with Inconsistent Auth
**File:** [app/api/admin_backend.py](app/api/admin_backend.py) and [app/api/endpoints/admin.py](app/api/endpoints/admin.py)
Two sets of module management endpoints are mounted to the same prefix. `admin_backend.py` uses API key auth (any valid module key can call it); `admin.py` uses JWT. This creates an auth bypass path.
**Fix:** Remove `admin_backend.py` endpoints. Consolidate to one set requiring JWT with an admin role check.
---
### H-5 — Exception Messages Returned to Clients
**File:** [app/api/endpoints/gemini.py](app/api/endpoints/gemini.py) lines 71–79, 142–143; similar in [app/api/endpoints/openai.py](app/api/endpoints/openai.py)
```python
except Exception as e:
return {"status": "error", "detail": f"Failed to parse text/plain as JSON: {str(e)}"}
```
Raw Python exception messages reveal framework internals, library versions, and data structures to callers.
**Fix:** Log the full exception server-side; return a generic message: `"Service error — please try again later"`.
---
### H-6 — No HTTPS Enforcement at Application Level
**File:** [docker-compose.yml](docker-compose.yml), [Caddyfile](Caddyfile)
TLS is terminated by Caddy, but the application itself has no HTTPS redirect or HSTS header. If Caddy is bypassed (direct port 8000 access), credentials are transmitted in plaintext.
**Fix:** Add an HTTPS enforcement middleware and set `Strict-Transport-Security: max-age=31536000; includeSubDomains` in all responses.
---
### H-7 — Unvalidated Query Parameter Casting
**File:** [app/api/endpoints/gemini.py](app/api/endpoints/gemini.py) lines 46–61
```python
temperature=float(params.get("temperature", 0.7)),
top_k=int(params.get("top_k", 40)),
```
Invalid input raises `ValueError` → 500. No bounds checking: `temperature` could be `-999`, `top_k` could be `2147483647`.
**Fix:** Use Pydantic `Query()` with `Field(ge=0.0, le=2.0)` constraints and handle `ValueError` with a 422 response.
---
### H-8 — Prompt Injection via User-Controlled System Prompt
**File:** [app/api/endpoints/gemini.py](app/api/endpoints/gemini.py) lines 110–112
```python
system_instruction = chat_data.system_prompt or GEMINI_SYSTEM_PROMPT
system_instruction += f"\n\nKnowledge Base:\n{chat_data.knowledge_base}"
```
Callers can fully replace the system prompt and append arbitrary content to it. This enables jailbreaking the model, data exfiltration via the LLM, and policy bypass.
**Fix:** Restrict `system_prompt` to an allowlist of predefined values. Never concatenate raw user input into a system instruction.
---
### H-9 — Stored XSS in Admin Panel (Module Names)
**File:** [app/static/admin.html](app/static/admin.html) line 608
```javascript
${m.name}
```
Module names from the API are interpolated directly into `innerHTML` via template literals. An attacker with a valid API key can create a module named `` to exfiltrate the admin JWT from `localStorage`.
**Fix:** Use `element.textContent = m.name` or DOMPurify. Add a `Content-Security-Policy` header.
---
### H-10 — XSS via Inline Event Handler with API Key
**File:** [app/static/admin.html](app/static/admin.html) lines 617, 620
```javascript