Circuit Breaker Labs builds crash test dummies for mental health chatbots

Circuit Breaker Labs builds crash test dummies for mental health chatbots

The self-funded startup uses simulated adversarial users to find where AI fails people in distress before real users do

Most AI doom talk focuses on a hypothetical future where machines turn on humanity. Circuit Breaker Labs is focused on a more immediate problem: chatbots that have already let vulnerable people down.

The US-based startup has built automated testing agents it describes as “crash test dummies” for conversational AI. The goal is simple to state and hard to execute: catch psychological harm before a real person in crisis ever types a message.

How the crash test works

Circuit Breaker Labs applies that logic to language models. Its core product is an agentic red-teaming framework, which is a fancy way of saying it deploys AI agents that play the role of difficult users.

These simulated users probe a chatbot for weak spots. The company targets three kinds of failure in particular.

The first is missing the cues of suicidal ideation, the signals that someone may be thinking about ending their life. The second is misreading slang, where a model takes a casual phrase literally or misses that a casual phrase carries serious meaning.

Advertisement

The third is subtler: gradual context manipulation. That is when a conversation drifts, one message at a time, until a model ends up somewhere it should never have gone.

To catch all three, the framework runs both single-turn and multi-turn tests. A single-turn test checks how a model responds to one prompt. A multi-turn test plays out a full conversation, which matters because harm often builds slowly rather than arriving in one obvious message.

By default, the testing suite centers on suicidal ideation scenarios. Teams can also build custom test groups and run repeated iterations to fit their own products.

Built for developers, not just ethicists

The company offers a command line interface called cbl, along with API access and GitHub Actions support. The GitHub piece is notable because it allows safety checks to run automatically every time a team ships new code.

The target customers are behavioral health AI teams, the builders working on emotional support and mental health applications.

The company also has clinical backing. It has received endorsements from clinical partners including Grow Therapy.

A launch date chosen with intent

Circuit Breaker Labs was founded in 2025. It made its public debut on September 10, 2025, which was World Suicide Prevention Day.

Co-founders Shirali Nigam and Arul Nigam have said their motivation came from earlier incidents in which AI chatbots failed to properly handle users in distress. The founders bring backgrounds in AI safety, psychology, and healthcare applications.

In February 2026, the company published a whitepaper laying out its methodology. The title leaned fully into the metaphor: “Crash-Test Dummies for AI: Agentic Red-Teaming for Mental Health Safety.”

As of mid-2026, Circuit Breaker Labs remains self-funded, with no reported funding rounds.

Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.
Circuit Breaker Labs builds crash test dummies for mental health chatbots
Circuit Breaker Labs builds crash test dummies for mental health chatbots

The self-funded startup uses simulated adversarial users to find where AI fails people in distress before real users do

Most AI doom talk focuses on a hypothetical future where machines turn on humanity. Circuit Breaker Labs is focused on a more immediate problem: chatbots that have already let vulnerable people down.

The US-based startup has built automated testing agents it describes as “crash test dummies” for conversational AI. The goal is simple to state and hard to execute: catch psychological harm before a real person in crisis ever types a message.

How the crash test works

Circuit Breaker Labs applies that logic to language models. Its core product is an agentic red-teaming framework, which is a fancy way of saying it deploys AI agents that play the role of difficult users.

These simulated users probe a chatbot for weak spots. The company targets three kinds of failure in particular.

The first is missing the cues of suicidal ideation, the signals that someone may be thinking about ending their life. The second is misreading slang, where a model takes a casual phrase literally or misses that a casual phrase carries serious meaning.

Advertisement

The third is subtler: gradual context manipulation. That is when a conversation drifts, one message at a time, until a model ends up somewhere it should never have gone.

To catch all three, the framework runs both single-turn and multi-turn tests. A single-turn test checks how a model responds to one prompt. A multi-turn test plays out a full conversation, which matters because harm often builds slowly rather than arriving in one obvious message.

By default, the testing suite centers on suicidal ideation scenarios. Teams can also build custom test groups and run repeated iterations to fit their own products.

Built for developers, not just ethicists

The company offers a command line interface called cbl, along with API access and GitHub Actions support. The GitHub piece is notable because it allows safety checks to run automatically every time a team ships new code.

The target customers are behavioral health AI teams, the builders working on emotional support and mental health applications.

The company also has clinical backing. It has received endorsements from clinical partners including Grow Therapy.

A launch date chosen with intent

Circuit Breaker Labs was founded in 2025. It made its public debut on September 10, 2025, which was World Suicide Prevention Day.

Co-founders Shirali Nigam and Arul Nigam have said their motivation came from earlier incidents in which AI chatbots failed to properly handle users in distress. The founders bring backgrounds in AI safety, psychology, and healthcare applications.

In February 2026, the company published a whitepaper laying out its methodology. The title leaned fully into the metaphor: “Crash-Test Dummies for AI: Agentic Red-Teaming for Mental Health Safety.”

As of mid-2026, Circuit Breaker Labs remains self-funded, with no reported funding rounds.

Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.