Circuit Breaker Labs tests AI systems for psychological safety
Circuit Breaker Labs is building an AI safety testing platform designed to find psychologically harmful chatbot behaviour before it reaches users. The five-person startup, founded by siblings Shirali Nigam and Arul Nigam, is a finalist in TechCrunch’s 2026 Startup Battlefield 200 and plans to pitch at TechCrunch Disrupt in San Francisco on October 13–15, 2026.
The company uses AI agents as simulated users, comparing them to an army of crash-test dummies. Its system runs from tens of thousands to hundreds of thousands of interactions each day, testing whether a model can recognise and respond safely to risky conversations that can unfold over time.
Testing language as people actually use it
Circuit Breaker Labs builds its simulations with human domain experts. The personas represent different ages, backgrounds, languages and cultures, and their conversations include slang, coded language, spelling mistakes and other features of ordinary communication.
Shirali Nigam, the company’s CEO, said models can handle standard speech patterns well but may fail when nuance varies between users. A child and an adult, a first- and second-language English speaker, or people using different forms of slang may phrase concerns in ways that produce very different model responses.
Those differences matter in sensitive settings. Arul Nigam, Circuit Breaker Labs’ CTO, said users, particularly younger people, can seek support from AI systems without receiving the help they need. The company’s focus is on vulnerabilities that arise during ordinary use rather than only when someone deliberately attempts to bypass safeguards.
Red-team evaluations for high-risk AI applications
The startup conducts red-team tests, an adversarial evaluation practice intended to expose weaknesses in AI models. Its testing is aimed at harmful psychological interactions, including situations where context is lost or a system misreads the meaning and implications of a user’s words.
Circuit Breaker Labs is currently operating as a safety testing lab for high-risk applications, including AI coaching, journaling and other mental-health support tools. It has a working product, although Arul Nigam did not identify the company’s marquee customers.
After each evaluation, the company applies a proprietary scoring method to produce scores it describes as auditable and explainable. The longer-term ambition is to test any application in which a person may form a parasocial relationship with a chatbot, including AI co-worker agents whose responses may differ from one conversation to another.
What deployment teams should take from the approach
For organisations developing conversational AI, the company’s approach highlights the importance of evaluating systems across sustained exchanges and realistic language variation. Safety testing for sensitive products should examine how models interpret context, slang and vulnerable users before those systems are put into day-to-day use.