mirror of
https://github.com/we-promise/sure.git
synced 2026-04-09 15:24:48 +00:00
* chore(helm): bump pipelock to v2.0.0 with trusted domains and redirect profiles - Bump pipelock image tag from 1.5.0 to 2.0.0 - Add first-class Helm values for trustedDomains and mcpToolPolicy.redirectProfiles - Update CI GitHub Action from @v1 to @v2 - Update compose example, config reference, and docs with v2.0 features * Releasing this today in `alpha` form --------- Co-authored-by: Juan José Mata <jjmata@jjmata.com>
223 lines
7.4 KiB
Markdown
223 lines
7.4 KiB
Markdown
# Pipelock: AI Agent Security Proxy
|
|
|
|
[Pipelock](https://github.com/luckyPipewrench/pipelock) is an optional security proxy that scans AI agent traffic flowing through Sure. It protects against secret exfiltration, prompt injection, and tool poisoning.
|
|
|
|
## What Pipelock does
|
|
|
|
Pipelock runs as a separate proxy service alongside Sure with two listeners:
|
|
|
|
| Listener | Port | Direction | What it scans |
|
|
|----------|------|-----------|---------------|
|
|
| Forward proxy | 8888 | Outbound (Sure to LLM) | DLP (secrets in prompts), response injection |
|
|
| MCP reverse proxy | 8889 | Inbound (agent to Sure /mcp) | Prompt injection, tool poisoning, DLP |
|
|
|
|
### Forward proxy (outbound)
|
|
|
|
When `HTTPS_PROXY=http://pipelock:8888` is set, outbound HTTPS requests from Faraday-based clients (like `ruby-openai`) are routed through Pipelock. It scans request bodies for leaked secrets and response bodies for prompt injection.
|
|
|
|
**Covered:** OpenAI API calls via ruby-openai (uses Faraday).
|
|
**Not covered:** SimpleFIN, Coinbase, Plaid, or anything using Net::HTTP/HTTParty directly. These bypass `HTTPS_PROXY`.
|
|
|
|
### MCP reverse proxy (inbound)
|
|
|
|
External AI assistants that call Sure's `/mcp` endpoint should connect through Pipelock on port 8889 instead of directly to port 3000. Pipelock scans:
|
|
|
|
- Tool call arguments (DLP, shell obfuscation detection)
|
|
- Tool responses (injection payloads)
|
|
- Session binding (detects tool inventory manipulation)
|
|
- Tool call chains (multi-step attack patterns like recon then exfil)
|
|
|
|
## Docker Compose setup
|
|
|
|
The `compose.example.ai.yml` file includes Pipelock. To use it:
|
|
|
|
1. Download the compose file and Pipelock config:
|
|
```bash
|
|
curl -o compose.ai.yml https://raw.githubusercontent.com/we-promise/sure/main/compose.example.ai.yml
|
|
curl -o pipelock.example.yaml https://raw.githubusercontent.com/we-promise/sure/main/pipelock.example.yaml
|
|
```
|
|
|
|
2. Start the stack:
|
|
```bash
|
|
docker compose -f compose.ai.yml up -d
|
|
```
|
|
|
|
3. Verify Pipelock is healthy:
|
|
```bash
|
|
docker compose -f compose.ai.yml ps pipelock
|
|
# Should show "healthy"
|
|
```
|
|
|
|
### Connecting external AI agents
|
|
|
|
External agents should use the MCP reverse proxy port:
|
|
|
|
```text
|
|
http://your-server:8889
|
|
```
|
|
|
|
The agent must include the `MCP_API_TOKEN` as a Bearer token in requests. Set this in your `.env`:
|
|
|
|
```bash
|
|
MCP_API_TOKEN=generate-a-random-token
|
|
MCP_USER_EMAIL=your@email.com
|
|
```
|
|
|
|
### Running without Pipelock
|
|
|
|
To use `compose.example.ai.yml` without Pipelock, remove the `pipelock` service and its `depends_on` entries from `web` and `worker`, then unset the proxy env vars (`HTTPS_PROXY`, `HTTP_PROXY`).
|
|
|
|
Or use the standard `compose.example.yml` which does not include Pipelock.
|
|
|
|
## Helm (Kubernetes) setup
|
|
|
|
Enable Pipelock in your Helm values:
|
|
|
|
```yaml
|
|
pipelock:
|
|
enabled: true
|
|
image:
|
|
tag: "2.0.0"
|
|
mode: balanced
|
|
```
|
|
|
|
This creates a separate Deployment, Service, and ConfigMap. The chart auto-injects `HTTPS_PROXY`/`HTTP_PROXY`/`NO_PROXY` into web and worker pods.
|
|
|
|
v2.0 adds trusted domain allowlisting, MCP tool redirect profiles, enhanced tool poisoning detection (full JSON schema scanning), and per-read kill switch preemption on long-lived connections. Process sandboxing and attack simulation are also available via `extraConfig` and CLI.
|
|
|
|
### Exposing MCP to external agents (Kubernetes)
|
|
|
|
In Kubernetes, external agents cannot reach the MCP port by default. Enable the Pipelock Ingress:
|
|
|
|
```yaml
|
|
pipelock:
|
|
enabled: true
|
|
ingress:
|
|
enabled: true
|
|
className: nginx
|
|
hosts:
|
|
- host: pipelock.example.com
|
|
paths:
|
|
- path: /
|
|
pathType: Prefix
|
|
tls:
|
|
- hosts: [pipelock.example.com]
|
|
secretName: pipelock-tls
|
|
```
|
|
|
|
Or port-forward for testing:
|
|
|
|
```bash
|
|
kubectl port-forward svc/sure-pipelock 8889:8889 -n sure
|
|
```
|
|
|
|
### Monitoring
|
|
|
|
Enable the ServiceMonitor for Prometheus scraping:
|
|
|
|
```yaml
|
|
pipelock:
|
|
serviceMonitor:
|
|
enabled: true
|
|
interval: 30s
|
|
additionalLabels:
|
|
release: prometheus
|
|
```
|
|
|
|
Metrics are available at `/metrics` on the forward proxy port (8888).
|
|
|
|
### Eviction protection
|
|
|
|
For production, enable the PodDisruptionBudget:
|
|
|
|
```yaml
|
|
pipelock:
|
|
pdb:
|
|
enabled: true
|
|
maxUnavailable: 1
|
|
```
|
|
|
|
See the [Helm chart README](../../charts/sure/README.md#pipelock-ai-agent-security-proxy) for all configuration options.
|
|
|
|
## Pipelock configuration file
|
|
|
|
The `pipelock.example.yaml` file (Docker Compose) or ConfigMap (Helm) controls scanning behavior. Key sections:
|
|
|
|
| Section | What it controls |
|
|
|---------|-----------------|
|
|
| `mode` | `strict` (block threats), `balanced` (warn + block critical), `audit` (log only) |
|
|
| `trusted_domains` | Allow internal services whose public DNS resolves to private IPs |
|
|
| `forward_proxy` | Outbound HTTPS scanning (tunnel timeouts, idle timeouts) |
|
|
| `dlp` | Data loss prevention (scan env vars, built-in patterns) |
|
|
| `response_scanning` | Scan LLM responses for prompt injection |
|
|
| `mcp_input_scanning` | Scan inbound MCP requests |
|
|
| `mcp_tool_scanning` | Validate tool calls, detect drift |
|
|
| `mcp_tool_policy` | Pre-execution rules, shell obfuscation, redirect profiles |
|
|
| `mcp_session_binding` | Pin tool inventory, detect manipulation |
|
|
| `tool_chain_detection` | Multi-step attack patterns |
|
|
| `websocket_proxy` | WebSocket frame scanning (disabled by default) |
|
|
| `logging` | Output format (json/text), verbosity |
|
|
|
|
For the Helm chart, most sections are configurable via `values.yaml`. For additional sections not covered by structured values (session profiling, data budgets, kill switch, sandbox, reverse proxy, adaptive enforcement), use the `extraConfig` escape hatch:
|
|
|
|
```yaml
|
|
pipelock:
|
|
extraConfig:
|
|
session_profiling:
|
|
enabled: true
|
|
max_sessions: 1000
|
|
```
|
|
|
|
## Modes
|
|
|
|
| Mode | Behavior | Use case |
|
|
|------|----------|----------|
|
|
| `strict` | Block all detected threats | Production with sensitive data |
|
|
| `balanced` | Warn on low-severity, block on high-severity | Default; good for most deployments |
|
|
| `audit` | Log everything, block nothing | Initial rollout, testing |
|
|
|
|
Start with `audit` mode to see what Pipelock detects without blocking anything. Review the logs, then switch to `balanced` or `strict`.
|
|
|
|
## Limitations
|
|
|
|
- Forward proxy only covers Faraday-based HTTP clients. Net::HTTP, HTTParty, and other libraries ignore `HTTPS_PROXY`.
|
|
- Docker Compose has no egress network policies. The `/mcp` endpoint on port 3000 is still reachable directly (auth token required). For enforcement, use Kubernetes NetworkPolicies.
|
|
- Pipelock scans text content. Binary payloads (images, file uploads) are passed through by default.
|
|
|
|
## Troubleshooting
|
|
|
|
### Pipelock container not starting
|
|
|
|
Check the config file is mounted correctly:
|
|
```bash
|
|
docker compose -f compose.ai.yml logs pipelock
|
|
```
|
|
|
|
Common issues:
|
|
- Missing `pipelock.example.yaml` file
|
|
- YAML syntax errors in config
|
|
- Port conflicts (8888 or 8889 already in use)
|
|
|
|
### LLM calls failing with proxy errors
|
|
|
|
If AI chat stops working after enabling Pipelock:
|
|
```bash
|
|
# Check Pipelock logs for blocked requests
|
|
docker compose -f compose.ai.yml logs pipelock --tail=50
|
|
```
|
|
|
|
If requests are being incorrectly blocked, switch to `audit` mode in the config file and restart:
|
|
```yaml
|
|
mode: audit
|
|
```
|
|
|
|
### MCP requests not reaching Sure
|
|
|
|
Verify the MCP upstream is configured correctly:
|
|
```bash
|
|
# Test from inside the Pipelock container
|
|
docker compose -f compose.ai.yml exec pipelock /pipelock healthcheck --addr 127.0.0.1:8888
|
|
```
|
|
|
|
Check that `MCP_API_TOKEN` and `MCP_USER_EMAIL` are set in your `.env` file and that the email matches an existing Sure user.
|