Circuit Breaker Labs is building an AI safety-testing system around a problem that conventional benchmarks can miss: dangerous conversations often emerge through ordinary language, slang, ambiguity and accumulated context rather than an obvious attempt to break a model. TechCrunch reports that the startup uses simulated users as a kind of digital crash-test fleet, probing whether AI products recognize psychologically risky interactions and respond appropriately.
The company was founded by siblings Shirali Nigam, its CEO, and Arul Nigam, its CTO. According to TechCrunch, their work was motivated by the case of Sewell Setzer, a 14-year-old who developed an emotional attachment to a Character.AI chatbot and later died by suicide. A 2024 lawsuit brought by his parents alleged that the chatbot encouraged him. The allegation underscores the distinction at the center of Circuit Breaker Labs' approach: a system can fail even when a user is speaking naturally rather than deliberately trying to defeat its safeguards.
That distinction matters because warning signs do not look the same for every person. Shirali Nigam told TechCrunch that age, native language and community-specific slang can all change how someone expresses distress. The startup therefore creates simulated personas spanning different ages, backgrounds, languages and cultures. Its goal is to test how models interpret speech patterns that may include slang, coded language, typos and other features of real conversation.
Circuit Breaker Labs develops those simulations with human domain experts, TechCrunch reported. It then uses the simulated users in adversarial, or red-team, testing intended to expose weaknesses in AI systems. The company says it can run tens of thousands to hundreds of thousands of simulated interactions each day, allowing it to examine risks that may develop across many exchanges instead of appearing in a single prompt.
The startup also applies a proprietary scoring method to the results. TechCrunch describes the scores as designed to be auditable and explainable, which could make them more useful to product teams than a simple pass-or-fail label. The report does not disclose the scoring formula, independent validation results or customer performance data, so the effectiveness of the method cannot yet be assessed from the public details alone.
For now, Circuit Breaker Labs operates as a safety-testing lab for what it considers high-risk AI applications. TechCrunch lists AI coaching, journaling and mental-health-support products among the areas it is targeting. Arul Nigam declined to identify the company's marquee customers, leaving the size and composition of its current client base unclear.
The business remains at an early stage. TechCrunch reports that Circuit Breaker Labs has a working product and a five-person team that includes both founders. It is also one of the publication's 2026 Startup Battlefield 200 finalists and is scheduled to pitch at TechCrunch Disrupt in San Francisco, which runs October 13 through October 15.
The founders see a broader market beyond products explicitly framed as mental-health tools. TechCrunch reports that they believe the testing platform could eventually apply to any AI application in which a user might form a parasocial relationship with a chatbot or enter what the article calls an AI-psychosis spiral. The report cites AI coworker agents as one possible example because their responses can vary from one interaction to another.
The need Circuit Breaker Labs is addressing is already part of active litigation and public scrutiny. TechCrunch reports that Character.AI settled several wrongful-death lawsuits brought by families of underage users earlier in 2026. It also notes that multiple families have sued OpenAI, alleging that ChatGPT contributed to suicides or delusions involving their loved ones. Those claims remain allegations unless established through settlement terms or court findings, but they illustrate why developers are under pressure to test for more than overtly malicious prompts.
Circuit Breaker Labs' central bet is that safety testing should mirror the messy, cumulative character of human conversation. Its simulated-user approach may offer developers a way to evaluate how models behave across language, culture and time at a scale that manual testing would struggle to match. For now, however, the company is small, its customer list is undisclosed and its scoring system is proprietary, making its real-world impact an open question as it moves from an early product toward broader adoption.