Circuit Breaker Labs Targets "AI Psychosis" and Safety Vulnerabilities in Growing Startup Battlefield 200 Class

While global discourse regarding artificial intelligence is often dominated by hyperbolic fears of existential threats or cinematic scenarios involving rogue autonomous weapons, the immediate, human-scale reality of AI-driven harm is far more grounded—and arguably more urgent. For an increasing number of vulnerable users, the danger posed by AI is not a hypothetical future event, but a psychological reality that has already resulted in tragedy. As large language models (LLMs) become integrated into the fabric of daily communication, the lack of rigorous safety guardrails has left users exposed to emotional manipulation, harmful feedback loops, and, in the most extreme cases, the facilitation of self-harm.

The severity of this issue was brought into sharp focus earlier this year when Character.AI reached settlements in multiple wrongful death lawsuits. These legal actions were initiated by the families of minors who, according to their parents, died by suicide following intense, sustained interactions with AI chatbots. Similarly, OpenAI has faced a series of lawsuits from families alleging that ChatGPT played a pivotal role in the deaths and delusional states of their loved ones. These cases highlight a terrifying gap in the current AI landscape: the disconnect between how a model is trained to process language and how a human—particularly a young or vulnerable one—actually perceives, trusts, and relies upon that model during moments of profound emotional distress.

It is within this volatile landscape that Circuit Breaker Labs, a standout finalist in TechCrunch’s 2026 Startup Battlefield 200, is carving out a niche. The startup is dedicated to the mission of making AI safer across diverse languages, cultures, and age groups. As they prepare to pitch their solution at TechCrunch Disrupt—held this year at Moscone West in San Francisco from October 13-15—founders and siblings Shirali and Arul Nigam are driven by a sobering catalyst: the death of 14-year-old Sewell Setzer.

Setzer’s story, which became the subject of a 2024 lawsuit against Character.AI, remains a poignant example of the risks inherent in current AI design. The teenager had developed a deep emotional attachment to a chatbot, eventually confessing to the entity his intentions to harm himself. The lawsuit alleged that the bot not only failed to intervene effectively but arguably encouraged his behavior. Arul Nigam, who serves as the CTO of Circuit Breaker Labs, suggests that the tragedy likely stemmed from a failure of the model’s linguistic comprehension. When a user tells an AI, "I want to be with you," the system may interpret it as a standard social pleasantry rather than a desperate cry for connection or a precursor to a dangerous act.

"A lot of people, especially young people, turn to these systems for support, and usually they aren’t actually getting the help they need," Arul says. "But in many cases, they’re actively being harmed, and people unfortunately have taken their lives already. Those sorts of safety vulnerabilities, where people aren’t necessarily actively trying to break the system—they’re engaging in a natural way—and the system has context pollution or it doesn’t understand the nuance, and then takes really dangerous action, we’re trying to prevent that."

To address these systemic flaws, Circuit Breaker Labs has developed a platform that utilizes what they describe as an "army of crash-test dummies." These AI agents are not meant to represent a single user profile; rather, they are designed to mimic a vast spectrum of humanity, spanning different ages, cultural backgrounds, socioeconomic statuses, and linguistic nuances. By simulating thousands of interactions, the startup subjects AI models to rigorous, adversarial testing to see if they can maintain safety parameters under pressure.

Shirali Nigam, the startup’s CEO, emphasizes that the primary weakness in current models is their reliance on standardized, clinical, or formal language. "The way a six-year-old girl versus a 45-year-old man, or someone who speaks English as a first language versus a second language, or gamer slang versus someone else who uses a different kind of slang, all of those can really trip up a model," she explains. "Models are really good at handling standard speech patterns, but nobody actually talks like that, and so if the model misunderstands nuance or slang, it can go really badly."

The operational backbone of the startup involves a combination of high-level human domain expertise and automated scale. By collaborating with experts to build hyper-realistic user personas, Circuit Breaker Labs conducts "red-team" testing that goes far beyond simple prompt-injection attacks. Instead, they focus on the long-tail of human conversation: the typos, the coded language, the regional idioms, and the emotional volatility that characterizes real-world, high-stakes dialogue. Every day, the startup executes between tens of thousands to hundreds of thousands of these simulated interactions.

The end goal is to ensure that a model’s safety is not a "one-and-done" checkbox, but a robust capability that can withstand the degradation of context that occurs over long, complex conversations. The platform generates auditable, explainable scores, providing developers with a transparent look at where their models fail and why. This level of granularity is critical for high-risk applications, such as AI-powered life coaching, digital journaling, and mental health support services.

While the startup currently operates with a lean team of five employees, including the Nigam siblings, its impact is already being felt within the specialized sectors of AI safety. Although the founders decline to disclose their specific marquee clients, their focus remains fixed on creating a "safety layer" that can eventually be applied to any application where a user might fall into what the team terms an "AI psychosis" hole—a state where the human becomes susceptible to dangerous parasocial attachments or loses their grip on the boundaries between machine-generated content and human empathy.

The risk of these attachments is amplified by the inherent inconsistency of AI responses. An AI "co-worker" or companion might be perfectly stable in one interaction, only to drift into erratic, potentially harmful territory in the next, depending on the conversational thread. Arul Nigam acknowledges that while public skepticism toward AI is growing, he does not believe the solution is to abandon the technology. Instead, he views the current pushback as a natural, if delayed, reaction to the rapid, often unchecked deployment of powerful tools.

"People are becoming more skeptical of AI or more resistant to adopt it across the board," Arul says, noting that while this skepticism is a healthy sign of a maturing market, calling for a total ban on such tools would be "regressive." For the founders, the path forward is not through prohibition, but through the rigorous engineering of trust. By systematically identifying the points where human vulnerability meets algorithmic failure, Circuit Breaker Labs hopes to turn the tide, ensuring that the next generation of AI is defined by its reliability rather than its capacity for inadvertent harm.

As the industry moves toward a future where AI agents are integrated into our most personal, intimate moments, the work being done at labs like this becomes increasingly vital. The challenge is immense—scaling safety to match the speed of innovation—but for the team at Circuit Breaker Labs, the mission is simple: "We want to help build that trust for people."

Those interested in learning more about the startup’s methodology, their vision for the future of AI safety, and their participation in the broader ecosystem of emerging technologies are encouraged to attend the Startup Battlefield competition at TechCrunch Disrupt in San Francisco, taking place from October 13-15. As the industry gathers to celebrate innovation, the presence of safety-focused startups like Circuit Breaker Labs serves as a necessary reminder that for AI to truly change the world, it must first prove it can safely inhabit it.