VC pitch
Encrypted guardrail interceptor for frontier agents.
Local proxy. Encrypted policy checks. Managed enterprise rollout through existing IT controls.
The problem
As frontier agents move from chat to execution, guardrails stop being policy docs and become runtime security controls. If the agent can read the rules, it can learn to work around them.
Adversarial prompt asks the agent to reveal or infer the hidden policy.
The agent leaks rules, categories, or threshold behavior.
The next prompt is optimized around the exposed control.
Thesis
Policy content is operational knowledge. Once exposed, it can be probed, inferred, and optimized against.
Admins should author and update guardrails without routing rule text through the same agent surface they are trying to constrain.
The workflow should stay the same until a request crosses the policy boundary.
Solution
CurtainWall intercepts LLM API calls, embeds the prompt locally, runs encrypted similarity against policy vectors, and blocks violations without revealing the guardrail corpus to the agent.
Security model
The proxy protects developer prompts from central collection. The Vault protects guardrail content from the agent. The agent gets a decision, not the rules.
| Component | Sees prompts | Sees guardrails | Sees scores | Role |
|---|---|---|---|---|
| Local proxy | Yes, plaintext | No, encrypted vectors only | No, encrypted scores only | Intercept and evaluate locally |
| Vault | Never | Embeddings only | Yes, decrypts | Judge, sync, RBAC, policy store |
| Agent | Its own conversation | Never | Never | Receives pass or refusal |
Product
Redirect existing Anthropic, OpenAI, and Google endpoints through localhost.
Security teams author guardrails through a web console served by the Vault, outside the agent loop.
Local proxy runs per workstation, avoiding a central plaintext prompt collector and fitting existing MDM controls.
Why us
Guardrail similarity runs over encrypted vectors using CryptoLab's HEaaN and EVI integration path.
The proxy model works across model providers and agent clients via base URL overrides.
Local daemon, Vault, RBAC, MDM rollout, and network hardening map to a real IT buying motion.
Go to market
The first buyer is the team already accountable for AI governance, secure SDLC, endpoint controls, and model access. The first wedge is local agent traffic from coding assistants and internal automation agents.
CISO, AI platform, DevSecOps, and platform engineering leaders.
Teams using Claude Code, Codex CLI, Gemini CLI, Cursor, or custom MCP-based agents.
From developer agents to CI, internal copilots, data workflows, and regulated automation.
Execution plan
Local proxy, Vault service, provider parsers, encrypted similarity adapter, RBAC, and tests.
Admin console, deployment packages, observability, network hardening guides, self-hosted review verification, and false-positive tuning.
Windows service support, SIEM integration, audit reports, and policy lifecycle workflows.
The ask
Priority: AI platform teams, security teams, and developer organizations already rolling out agentic coding or internal automation.
Funding goes toward package distribution, admin console, deployment hardening, auditability, and customer pilots.
CurtainWall is proprietary enterprise software by CryptoLab Inc.