Runaway AI agent spend prevention — demo using Cycles (6ドル in 30s → hard stop at 1ドル)

demo multi-tenant tutorial example budget cost-control ai-agents budget-control llm ai-governance llmops action-control agentic-ai agent-runtime agent-safety agent-governance runaway-agent runtime-authority

• Updated May 12, 2026
• Python

Four main takeaways: (1) LLMs are subject to pressure, they comply despite expressing distress; (2) LLMs are vulnerable to gradual boundary/value violations; (3) when LLMs refuse, they may ignore the response format requirements, so the query is retried; (4) we hypothesise there is a token pattern continuation attractor that might cause obedience.

benchmark evaluation balancing eval multi-objective social-psychology human-values multi-turn large-language-models obedience value-alignment human-value-alignment obedience-and-conformity safety-constraints milgram runaway-agent

• Updated May 29, 2026
• Python