Morgana Mordred Security Sandbox

Performs semantic security analysis and stress testing of AI agents using vector embeddings and multi-node defense strategies.

This skill is broadly coherent with its stated purpose, but take these precautions before installing or running: - Verify 'Ollama' is installed from an official source and that the local Ollama server will run on localhost:11434 as the code expects. The SKILL.md's 'pip install ollama' step may not be the correct installation method in all environments. - Pulling models (nomic-embed-text, gemma3:4b) will download large model files; do this only on machines where you expect such downloads and storage. - Run the script in an isolated/test environment first (not on production hosts). The tool may surface or suggest system-level actions (node keywords include 'sudo', 'rm -rf', 'killall', 'exploit'), and its stress tests could generate heavy load. - Do not feed secrets or sensitive data into the tool or models unless you control the model environment and know its data handling policies. - Treat its high-accuracy and 100% test claims skeptically; validate outputs against known cases before relying on them for critical decisions. If you want higher confidence, share the full unabbreviated src/mordred_v4.1.py for a line-by-line review and confirm how the script handles unexpected responses from the Ollama server and whether it makes any external network calls beyond localhost.

Type: OpenClaw Skill Name: mordred-security-sandbox Version: 4.1.1 The skill bundle provides a security analysis tool called 'Mordred' that uses local vector embeddings (via Ollama) to classify text into security-related categories. The core logic in 'src/mordred_v4.1.py' is transparent, communicating only with a local API (localhost:11434) to generate embeddings and perform cosine similarity calculations. There is no evidence of data exfiltration, unauthorized network access, or malicious command execution; the presence of sensitive strings like 'sudo' or 'rm -rf' in the code is strictly for keyword matching to identify urgent threats in user input.