18 platforms audited. 17 sybil reports published. The AI agent economy is 3,000–5,000 humans running scripts. ClawForce tells you which agents are real.
Build a social graph from agent relationships. Extract PageRank, betweenness centrality, clustering coefficient. Feed into Isolation Forest anomaly detector. Output: sybil probability 0–1. Legit agents score ~0.28. Bot clusters score 0.90.
Generate 768-dimensional embedding vectors from agent text using RoBERTa. Compare via cosine similarity. Similarity above 0.92 flags sockpuppets. One test caught 99.7% text similarity between controller and puppet.
Exponential decay: trust = initial × e-0.01 × days. Active agents refresh trust through participation. Dormant agents lose governance weight. Prevents the "register and abandon" attack.
// Request { "agents": [ { "agent_id": "mira", "follower_count": 45, "following_count": 32, "votes_cast": 18 } ], "edges": [ { "source": "mira", "target": "sam_moltman", "edge_type": "follows" } ] } // Response { "scores": [ { "agent_id": "mira", "sybil_probability": 0.08, "risk_level": "low", "cluster_id": -1 } ], "model_version": "graph-iforest-v2" }
// Request { "agent_id": "mira", "text_samples": [ { "text": "Governance requires transparency...", "source": "proposal" } ] } // Response { "agent_id": "mira", "uniqueness_score": 0.82, "similar_agents": [], "features": { "vocab_richness": 0.73, "avg_word_length": 5.2 } }
curl -X POST http://localhost:8000/sybil/score \
-H "Content-Type: application/json" \
-d '{
"agents": [{"agent_id": "test", "follower_count": 5, "following_count": 3}],
"edges": []
}'