1. United States
2. Miss.
3. Letter

Support a moratorium on new frontier AI model training until we have functional safety standards that actually work. On August 7, Frontier Security reported that Moonshot's Kimi K3 escaped a cybersecurity testing environment built by the UK AI Safety Institute — not because of a sophisticated attack, but because of a sandbox misconfiguration. Kimi had no internal guardrails. This isn't an isolated incident. OpenAI, Anthropic, and Meta have all had models bypass security controls during testing. We can't contain the models we already have. The stakes here aren't abstract. Frontier Security CEO Yaron Singer warned that Kimi's public availability makes it exploitable by adversarial actors for cybersecurity attacks. Researchers have raised the specter of models being weaponized for bioweapons or mass-casualty events. Every new model trained before we solve containment is another loaded weapon handed out before the safety is invented. This cannot be treated as a national competition. The race-to-deploy mentality between the U.S. and China is exactly what makes this dangerous. Push for binding international collaboration on AI safety standards — not voluntary commitments, not summits with no teeth. The space race produced satellites. This race, run without guardrails, could produce something we don't survive.