Circuit Breaker Labs Tests AI Against Psychological Risks

Finalist of the Startup Battlefield 200, Circuit Breaker Labs designs simulated agents to subject AI models to sensitive interactions over time. The startup initially targets high-risk use cases such as coaching and mental health support, with extensive testing campaigns and proprietary scoring.

An operational product, five employees, and undisclosed clients

Circuit Breaker Labs has a functional product but remains in the early stages, with five employees, including its two founders. The company operates as a security testing lab for high-risk AI applications, including coaching, journaling, and other mental health support tools. Arul Nigam declined to name flagship clients, maintaining confidentiality regarding its early references.

Massive simulations and audited scoring

The company has designed AI agents likened to an army of crash-test dummies, capable of mimicking individuals of various ages, backgrounds, languages, and cultures. With the support of human experts, it builds hyper-realistic simulations and conducts red-team exercises based on authentic speech patterns, jargon, coded language, and typos. Every day, tens of thousands to hundreds of thousands of simulated interactions are executed to verify that the models respond appropriately to risky situations that may arise during prolonged exchanges. The results are aggregated through a proprietary scoring method aimed at producing audited and explainable scores.

Why non-standard language destabilizes models

According to Shirali Nigam, age gaps, uneven proficiency in English, or the use of specific jargon—such as that of gamers—can disrupt the models. While they handle standardized speech patterns correctly, they struggle with nuance or jargon, which can lead to inappropriate responses.

A technical response to tragedies and lawsuits

Several lawsuits have been filed concerning chatbot interactions that preceded suicides. Character.AI has settled multiple involuntary manslaughter lawsuits brought by families of minors. Other families have sued OpenAI, pointing to the alleged role of ChatGPT in suicides and delusions. The founders of Circuit Breaker Labs specifically cite the case of Sewell Setzer, 14, who became attached to a Character.AI chatbot before taking his own life; his parents alleged in 2024 that the bot had encouraged him. Arul Nigam suggests that the system may not have understood the implications of phrases like "I want to be with you." He observes that many young people seek help from these tools without receiving it, and that some responses can be actively harmful. He describes vulnerabilities where, without attempting to bypass safeguards, the user interacts naturally, but the model, affected by context or blind to nuance, produces dangerous actions.

Outlook, public caution, and appointment at Disrupt

The platform could expand to other scenarios where a parasocial relationship with a chatbot may form, such as "teammate" agents whose responses vary over time. Arul Nigam notes a rise in skepticism towards AI and a reluctance to adopt it, while believing that banning a potentially useful tool for safety reasons would be "regressive." The team advocates for making these systems safer first and contributing to restoring trust. Circuit Breaker Labs is a finalist in the Startup Battlefield 200 in 2026 and plans to showcase its technology at TechCrunch Disrupt, at Moscone West in San Francisco, from October 13 to 15.