Add secureagentics/Adrian

agents/secureagentics__adrian/README.md

@@ -0,0 +1,55 @@
+# Adrian — Runtime Security Monitor for AI Agents
+
+**Adrian** is an open-source, [AARM-aligned](https://aarm.dev) runtime security monitoring and control engine for AI agents. It watches every tool call, API interaction, and reasoning trace in real time — detecting malicious, misaligned, or out-of-remit behaviour before or as it happens.
+
+## What It Does
+
+Most agent monitors watch *what* an agent does (activity logs). Adrian also watches *why* — correlating an agent's reasoning traces with its actions to give ~35% better detection accuracy than behaviour-only approaches (per OpenAI / DeepMind research, 2025).
+
+**Key capabilities:**
+
+- 🔍 **Prompt injection detection** — catches attempts to hijack or redirect the agent mid-session
+- 🛠️ **Tool-call auditing** — every MCP call, API hit, DB write, and file operation is logged and classified
+- 🧠 **Reasoning-trace analysis** — understands *why* the agent took an action, not just *what* it did
+- 🚨 **Policy-drift detection** — flags when an agent starts operating outside its configured remit
+- 🛑 **In-flight intervention** — runs in *audit* (log + alert) or *block* (halt before execute) mode
+- 🏠 **Self-hostable** — full stack (Go backend + Next.js dashboard + Llama.cpp Gemma classifier) on a single Docker host, no external telemetry
+
+## Quick Install
+
+```python
+pip install adrian-sdk
+
+import adrian
+from langchain_openai import ChatOpenAI
+
+adrian.init(api_key="adr_live_...")
+llm = ChatOpenAI(model="gpt-4o")
+response = await llm.ainvoke("your agent prompt here")
+adrian.shutdown()
+```
+
+That's it — all LangChain / LangGraph calls are automatically instrumented. Events appear in the dashboard within seconds.
+
+## Self-Hosting
+
+Adrian ships a complete Docker Compose stack with a Go backend, Next.js dashboard, and a local Llama.cpp container running Gemma 4 (E2B / E4B). No cloud dependency. No telemetry leaving the box.
+
+```sh
+git clone https://github.com/secureagentics/Adrian
+cd Adrian
+docker compose --profile setup run --rm setup bootstrap
+docker compose --profile llm up -d
+```
+
+## Why Adrian?
+
+Most monitoring tools in this space are ML classifiers trained on labelled prompt-injection datasets. Adrian uses world-model reasoning: it holds a working understanding of what the agent is *supposed* to be doing and assesses each new action against that. If your e-commerce agent starts resetting passwords, that's a flag — even if it has never appeared in any training data.
+
+## Links
+
+- 📖 [Documentation](https://docs.adrian.secureagentics.ai)
+- 📊 [Managed Dashboard](https://app.adrian.secureagentics.ai)
+- 🐍 [PyPI — adrian-sdk](https://pypi.org/project/adrian-sdk/)
+- 💬 [Discord](https://discord.gg/6nmJ9k3u6)
+- 🔗 [GitHub](https://github.com/secureagentics/Adrian)

agents/secureagentics__adrian/metadata.json

@@ -0,0 +1,14 @@
+{
+  "name": "adrian",
+  "author": "secureagentics",
+  "description": "Open-source runtime security monitor for AI agents — detects prompt injection, tool-call abuse, and policy drift by analysing both behaviour and reasoning traces.",
+  "repository": "https://github.com/secureagentics/Adrian",
+  "version": "1.0.0",
+  "category": "security",
+  "tags": ["runtime-security", "prompt-injection", "agent-monitoring", "policy-drift", "tool-auditing", "langchain", "langgraph", "mcp", "aarm", "self-hosted"],
+  "license": "Apache-2.0",
+  "model": "anthropic:claude-sonnet-4-6",
+  "adapters": ["claude-code", "system-prompt"],
+  "icon": false,
+  "banner": false
+}