Lilith Zero is a high-performance security runtime that mitigates data exfiltration and unauthorized tool invocation in LLM-based agent systems. It interposes at the transport layer, enforcing security invariants through deterministic policy evaluation and strictly framed execution.

OS, framework, and language agnostic: uniform security primitives across all implementation environments.

Technical Fundamentals

• Fail-Closed Architecture: defaults to DENY. Missing policy, parse error, or internal fault → all traffic blocked.
• Zero-Trust Transport: stdio and network payloads treated as potentially malicious; strict Content-Length framing prevents JSON smuggling.
• Type-Safe Invariants: Rust type system makes invalid security states (e.g. unverified taint propagation) unrepresentable at compile time.
• Deterministic Classification: tool classes are declared explicitly in policy; no runtime heuristics.

Core Capabilities

System Architecture

graph LR
    Agent["AI Agent / LLM"] -->|stdio| Codec

    subgraph lz["Lilith Zero (Rust)"]
        Codec["Hardened Codec"] --> Coord["Session Coordinator"]
        Coord --> Policy["Policy Engine"]
        Coord --> Taint["Taint Tracker"]
        Coord --> Audit["Audit Logger"]
        Coord --> Proc["Process Supervisor"]
    end

    Proc -->|stdio| ToolLocal["MCP Server (local)"]
    Coord -.->|Streamable HTTP| ToolRemote["MCP Server (remote)"]

The agent always communicates with Lilith Zero over stdio: the transport setting only changes how Lilith Zero connects upstream: spawning a local child process (stdio mode) or acting as an HTTP client to a remote MCP server (Streamable HTTP mode). Policy enforcement, taint tracking, and audit logging apply in both cases.

Components

Agent Integrations

All adapters share the same policy engine, taint tracker, and audit layer: only the I/O contract differs.

Claude Code

Hooks fire on every tool call. Lilith reads the JSON event from stdin and signals via exit code: 0 = allow, 2 = block.

.claude/settings.json (project) or ~/.claude/settings.json:

{
  "hooks": {
    "PreToolUse":  [{"matcher": "", "hooks": [{"type": "command", "command": "lilith-zero hook --policy /path/to/policy.yaml"}]}],
    "PostToolUse": [{"matcher": "", "hooks": [{"type": "command", "command": "lilith-zero hook --policy /path/to/policy.yaml"}]}]
  }
}

See scripts/hooks.json for a ready-to-use template.

GitHub Copilot CLI

Copilot reads {"permissionDecision":"allow"|"deny"} from stdout; exit code is always 0.

~/.config/github-copilot/hooks.json:

{
  "preToolUse":  {"command": "lilith-zero hook --format copilot --event preToolUse --policy /path/to/policy.yaml"},
  "postToolUse": {"command": "lilith-zero hook --format copilot --event postToolUse --policy /path/to/policy.yaml"}
}

See examples/gh-copilot/ for complete policy examples (bash, view, rg, glob, report_intent).

VS Code Copilot Sidebar

Event type is inferred from the payload: no --event flag required.

.vscode/settings.json:

{
  "github.copilot.chat.agent.hooks": {
    "PreToolUse":  {"command": "lilith-zero hook --format vscode --policy /path/to/policy.yaml"},
    "PostToolUse": {"command": "lilith-zero hook --format vscode --policy /path/to/policy.yaml"}
  }
}

See examples/vscode/ (editFiles, runTerminalCommand, #fetch, #codebase, MCP tools).

OpenClaw

OpenClaw uses stdio MCP only. Wrap each MCP server via lilith-zero run in ~/.openclaw/openclaw.json:

{
  "mcp": {
    "servers": {
      "filesystem": {
        "command": "lilith-zero",
        "args": ["run",
                 "--upstream-cmd", "npx -y @modelcontextprotocol/server-filesystem /home/user/projects",
                 "--policy", "~/.lilith/openclaw-policy.yaml"],
        "transport": "stdio",
        "env": {"LILITH_ZERO_PIN_FILE": "~/.lilith/pins/filesystem.json"}
      }
    }
  }
}

See examples/openclaw/ for policy-base.yaml (conservative) and policy-paranoid.yaml (read-only).
The examples/openclaw/CVE-COVERAGE.md maps all 139 OpenClaw advisories to the corresponding lilith-zero controls.

Forward-looking: once OpenClaw ships its hook system (#60943), replace the run wrapper with:
lilith-zero hook --format openclaw --policy ~/.lilith/openclaw-policy.yaml

Copilot Studio Webhook

REST server implementing the Microsoft Copilot Studio External Threat Detection API.

# Dev: no auth
lilith-zero serve --bind 127.0.0.1:8080 --auth-mode none --policy policy.yaml

# Production: Microsoft Entra ID (RS256)
lilith-zero serve --bind 0.0.0.0:8443 \
  --auth-mode entra \
  --entra-tenant-id <TENANT_GUID> \
  --entra-audience https://security.contoso.com \
  --policy policy.yaml

Endpoints: POST /validate (health), POST /analyze-tool-execution (policy evaluation).

Session Persistence

Hook mode maintains taint state across binary restarts within a session. State is stored in ~/.lilith/sessions/<session-id>.json under a cross-process file lock (LockFileEx on Windows, flock on Unix). Taints accumulated in one tool call are visible to the next invocation.

Transports

stdio (default)

lilith-zero run --upstream-cmd "python mcp_server.py" --policy policy.yaml

Wraps a child MCP server process. Lilith Zero sits between agent and server on stdin/stdout.

Streamable HTTP (MCP 2025-11-25)

lilith-zero run --transport http --upstream-url http://localhost:9000/mcp --policy policy.yaml

Connects to a remote MCP server; no child process spawned.

SDKs

Python

uv add lilith-zero          # auto-downloads the correct binary for your platform
# or: pip install lilith-zero

from lilith_zero import Lilith

async with Lilith("python mcp_server.py", policy="policy.yaml") as lz:
    result = await lz.call_tool("read_file", {"path": "/data/report.txt"})

TypeScript / Node.js

bun add @badcompany/lilith-zero
# or: npm install @badcompany/lilith-zero

import { Lilith } from "@badcompany/lilith-zero";

await using lz = new Lilith({ upstream: "bun mcp_server.ts", policy: "policy.yaml" });
const result = await lz.callTool("read_file", { path: "/data/report.txt" });

Setup

1. Install

Python SDK (manages the binary automatically):

uv add lilith-zero
# or: pip install lilith-zero

Shell (Unix/macOS):

curl -sSfL https://badcompany.xyz/lilith-zero/install.sh | sh

PowerShell (Windows):

irm https://badcompany.xyz/lilith-zero/install.ps1 | iex

2. Write a policy

id: my-policy
customer_id: my-org
name: "My Agent Policy"
version: 1

protect_lethal_trifecta: true   # auto-blocks ACCESS_PRIVATE + UNTRUSTED_SOURCE → EXFILTRATION

# Declare tool classes explicitly: Lilith performs no heuristic inference.
tool_classes:
  curl:         [EXFILTRATION]
  send_email:   [EXFILTRATION]

static_rules:
  read_file:   ALLOW
  calculator:  ALLOW
  bash:        DENY

taint_rules:
  - tool: read_file
    action: ADD_TAINT
    tag: ACCESS_PRIVATE

  - tool: curl
    action: ADD_TAINT
    tag: UNTRUSTED_SOURCE

resource_rules: []

protect_lethal_trifecta: true injects a rule that blocks any tool declared as EXFILTRATION class when both ACCESS_PRIVATE and UNTRUSTED_SOURCE taints are simultaneously active. Tool class assignments are always explicit in the policy: no name-based inference.

3. Wire it to your agent

# Claude Code: one-liner setup
echo '{"hooks":{"PreToolUse":[{"matcher":"","hooks":[{"type":"command","command":"lilith-zero hook --policy $(pwd)/policy.yaml"}]}]}}' \
  > .claude/settings.json

4. Environment variables

Observability

Audit Logs

Every decision: ALLOW and DENY: emits a tamper-evident JSONL line to stderr:

[AUDIT] <HMAC-SHA256-base64url> {"session_id":"…","timestamp":…,"event_type":"Decision","details":{"decision":"DENY","tool_name":"curl"}}

Verifying the signature against the session key proves the log was not altered after the fact.

lilith-telemetry

lilith-telemetry is a companion sidecar that aggregates Lilith Zero security events into structured, distributed traces. It implements the TelemetryHook trait and routes spans through a configurable Flock dispatch pipeline.

# Run with telemetry enabled (Flock link from your dashboard)
lilith-zero-telemetry run \
  --upstream-cmd "python mcp_server.py" \
  --policy policy.yaml \
  --flock-link "lilith://collector:7700?key_id=0x<hex>"

Each tool call produces a parent span with child spans for codec framing, policy evaluation, and the upstream call. Decisions are tagged LEVEL: ROUTINE (allow) or LEVEL: CRITICAL (deny/violation):

SESSION: 189e9a3a…  TRACE: 189e9a44…  SPAN: 189e9a44…  LEVEL: CRITICAL  KIND: Server
MSG: Blocked tool web_search: Web search prohibited after database access (exfiltration prevention)

Multi-tool interactions from a single LLM reasoning step are grouped into unified trace spans, enabling precise audit of the full agent decision chain.

Security & Assurance

View full SECURITY.md for proof harnesses and audit logs.

Performance

Micro-benchmarks (core logic):

• Codec decoding: ~247 ns / message
• Policy evaluation: ~660 ns / rule

End-to-end (Windows, release build, p50):

• RPC overhead: < 0.5 ms
• Startup: ~15 ms (process spawn + handshake)
• Memory: ~4 MB RSS
• Throughput: > 1.5 M validations/sec

Development

See docs/development.md for full setup.

# Rust core (from lilith-zero/)
cargo build && cargo test --all-features
cargo clippy --all-targets --all-features -- -D warnings

# Python SDK (from sdk/)
uv venv && uv pip install -e ".[dev]" && uv run pytest tests -v

License

Apache License, Version 2.0. See LICENSE.