A five-person startup called Circuit Breaker Labs is trying to catch the ways AI companions fail vulnerable users before those failures reach real people, by unleashing armies of simulated users on chatbots to probe for psychological blind spots. The company, a Startup Battlefield 200 finalist heading into TechCrunch Disrupt this month, builds AI agents designed to act as "crash-test dummies" that converse with mental-health and companion chatbots the way real, often troubled, users do.
Founded by siblings Shirali Nigam, the CEO, and Arul Nigam, the CTO, the company runs what it calls agentic red-teaming: simulated conversations, run tens of thousands to hundreds of thousands of times a day, that vary by age, language, culture and speech pattern in search of the moments where a chatbot misreads context and responds in a way that could cause harm, according to TechCrunch's profile of the company. The startup has not disclosed funding, pricing or named customers, describing itself as still in its early stages.
Models are really good at handling standard speech patterns, but nobody actually talks like that.
Shirali Nigam, co-founder and CEO, Circuit Breaker Labs
A response to documented harm
The Nigams founded the company after reading about the case of Sewell Setzer, a 14-year-old who died by suicide after developing an emotional attachment to a Character.AI chatbot he had confided self-harm thoughts to. Character.AI has since settled wrongful-death lawsuits brought by families of users who died by suicide, and several families have separately sued OpenAI alleging ChatGPT played a role in relatives' deaths or mental-health crises. Those cases have turned what was once a hypothetical concern — that conversational AI could miss or even encourage expressions of distress — into an active legal and product-liability question for the industry.
Arul Nigam said the failures Circuit Breaker Labs looks for are rarely the result of someone deliberately trying to "jailbreak" a model. More often, he said, a user is "engaging in a natural way" and the model simply misunderstands the context — a misspelled word, a regional idiom or coded slang can be enough to slip past a safety guardrail that would catch a more plainly worded request.
The company argues that AI chatbot safety needs the kind of independent, pre-launch testing regime that exists for food and cars, rather than the after-the-fact scrutiny that currently follows a harm once it becomes public and lands in court. That pitch is arriving at a moment when regulators and plaintiffs' lawyers alike are paying closer attention to how AI companion apps are tested before they reach teenagers and other vulnerable users, giving a small testing-focused startup an opening that did not clearly exist even a year ago.
