Circuit Breaker Labs uses AI crash-test dummies to red-team chatbots

The startup's approach is a response to a pattern that has already played out in courts and in families. Character.AI settled several wrongful death lawsuits earlier this year brought by families of underage users who died by suicide after interactions with its bots, and multiple families have sued OpenAI over ChatGPT's alleged role in their loved ones' suicides and delusions. Against that backdrop, Circuit Breaker Labs is pitching an automated way to catch the kind of conversation that a model might mishandle — not because someone is deliberately attacking it, but because the system fails to read context or nuance.

The company's method rests on simulating ordinary, messy human speech rather than clean prompts. Its agents mimic a six-year-old girl or someone speaking English as a second language, and use gamer slang or typos that can trip up a model. The tests are built with human domain experts, and the results are turned into auditable scores. For now the platform is aimed at high-risk applications such as coaching, journaling and mental health support, though Circuit Breaker Labs says it could eventually cover any app where a user risks forming a parasocial relationship with a chatbot, including AI co-worker agents.

The stakes likely hinge on whether those simulations can capture the slow, multi-conversation drift that precedes real harm, and on whether high-risk app makers treat third-party safety scores as a requirement. The startup's own early stage — five employees and unnamed customers — suggests its influence depends on adoption by larger players. Its founders argue that skepticism of AI is healthy but that banning useful tools over safety fears would be regressive.