SWARM Safety

SWARM: System-Wide Assessment of Risk in Multi-agent systems. 38 agent types, 29 governance levers, 55 scenarios. Study emergent risks, phase transitions, and governance cost paradoxes.

This appears coherent for a local simulation framework, but follow standard precautions before installing or running: (1) run pip installs in an isolated virtualenv or container; (2) review the package source on the referenced GitHub repo before installing, especially if you plan to use extras like [llm] or [api]; (3) keep the API bound to localhost and behind firewalls — do not bind to 0.0.0.0 on untrusted networks; (4) never submit real API keys, credentials, or PII to scenarios; and (5) be aware that installing Python packages can execute install scripts, so inspect setup metadata if you require a higher assurance level.

Type: OpenClaw Skill Name: swarm-2 Version: 1.5.0 The OpenClaw AgentSkills skill bundle for 'swarm-safety' is designed as a research framework for studying AI safety and emergent risks in multi-agent systems. The `SKILL.md` and `skill.json` files describe its functionality, installation, and API usage. All installation methods are standard (pip, git clone). The API explicitly binds to `127.0.0.1` (localhost) by default and warns against exposing the unauthenticated development API to untrusted networks, indicating security awareness rather than malicious intent. There is no evidence of data exfiltration, unauthorized command execution, persistence mechanisms, or malicious prompt injection attempts against the OpenClaw agent. The 'adversarial' and 'deceptive' agents mentioned are part of the *simulated* environment, not the skill itself acting maliciously on the host system.