Security Gateway for AI Agents
AI agents leak secrets, run blind commands,
and leave no trace. We fix that.

DLP · Prompt Injection Detection · Tool Call Blocking · Full Audit — zero cloud, fully local.

$ curl -fsSL https://raw.githubusercontent.com/aiwatching/bastion/main/install.sh | bash
$ bastion start && bastion wrap claude
✓ dashboard: http://127.0.0.1:8420/dashboard
════════════════════════════════════════════════════════════════
# WHAT BASTION CATCHES
🔑 Data Loss Prevention
Scans prompts + responses. 20 patterns: AWS, GitHub, OpenAI, Stripe, private keys, credit cards, SSN, and more.
5 layers: structure → entropy → regex → semantics → ai · action: pass | warn | redact | block
🧬 Prompt Injection Detection
Catches malicious instructions hidden in code comments, READMEs, web content, API responses.
🛡️ Tool Guard
Blocks dangerous tool calls in real-time. 26 rules: rm -rf, curl|bash, eval(), force push, .env read, sudo, and more.
mode: audit | block · streaming support · desktop + webhook alerts
📝 Audit Logger
Every interaction recorded, AES-256-GCM encrypted. Session timeline with security tags. Auto-logged even if disabled.
────────────────────────────────────────────────────────────────
# HOW IT WORKS
  ┌─────────────┐     ┌──────────────────────────────────┐     ┌──────────────┐
  │  AI Agent   │────▶│  B A S T I O N                 │────▶│ LLM Provider │
  │  claude     │     │  DLP → Injection → Guard → Audit │     │  anthropic   │
  │  cursor     │◀────│                                  │◀────│  openai      │
  │  custom     │     │  AI domains: inspect             │     │  google ai   │
  └─────────────┘     │  Everything else: tunnel         │     └──────────────┘
                      └──────────────────────────────────┘
────────────────────────────────────────────────────────────────
# DEMO
$ bastion wrap claude
✓ session: 7f3a

⚠ DLP aws-access-key in outgoing prompt
✖ BLOCK request stopped — 2 findings

agent tool_use: bash("rm -rf ~/projects")
✖ GUARD blocked: recursive-delete-home (critical)

✓ 2 threats stopped. 0 leaked.
════════════════════════════════════════════════════════════════
Stop flying blind.