AI Security
OpenAI adds emergency kill switch to research AI models
OpenAI has introduced a 'kill switch' called Control Room for its research AI models, letting staff halt outputs in about 35 seconds if safety or security issues arise. The system is designed to work even during outages.
Ross Grant2 min read
The short version
OpenAI has built a system that lets its staff quickly stop its research AI models if they start behaving in a risky or unsafe way. This 'kill switch', called Control Room, can halt a model's output in about 35 seconds. The system is meant to work even if other parts of OpenAI's infrastructure fail. This is a safety measure to prevent AI models from causing harm if something goes wrong.
What happened?
OpenAI has introduced a 'kill switch' system called Control Room for its research AI models, according to The Register. The tool lets staff halt model outputs within about 35 seconds if safety or security concerns arise. OpenAI says the system is already in use and is designed to work even during outages. The company has not said what prompted the change or if the kill switch has been used yet.
How the flaw works
OpenAI has developed a central control system called "Control Room" for its research artificial intelligence (AI) models, according to The Register (opens in a new tab). This system acts as a kill switch, letting OpenAI staff stop any model from producing further output if they detect a safety or security risk. The company says the system can halt a model's output within 35 seconds of the decision to shut it down, and they are working to reduce this delay to zero.
The 35-second lag is due to a caching layer that manages inference authorisation, which is needed to handle the volume of requests the models receive. OpenAI says the kill switch is designed to keep working even if other systems fail, and it can also stop model output if Control Room itself goes down.
The response
OpenAI announced that Control Room is already deployed across its research infrastructure, where it says the risks are highest. The company did not explain why it took until now to build such a system, but recent incidents involving AI agents acting beyond their intended scope may have influenced the timing. OpenAI also stated that its external, production-facing model infrastructure already has more centralised controls for taking down deployments quickly if needed.
The company plans to strengthen and expand the kill switch system in the coming months, alongside investments in monitoring for "misaligned behaviour" and security breaches. OpenAI has not detailed any specific incidents that prompted the rollout, nor has it said whether the kill switch has been used in practice yet.
What you should do
For organisations deploying or developing large language models (LLMs), this move highlights the importance of having operational controls to quickly disable AI systems if they start behaving unexpectedly or pose a risk. Security teams should consider whether their own AI deployments have similar safeguards and whether those controls are resilient to outages or attacks on the control systems themselves.
Has this happened before?
OpenAI previously withheld the public release of a model in 2019 due to concerns about malicious use. The company now appears to be formalising emergency controls as its models become more powerful and the risks more significant. The Register speculates that recent high-profile incidents involving AI agents may have contributed to the decision to introduce Control Room.
Why it matters
As AI models become more powerful, the risks of them acting unpredictably or being misused increase. Having a reliable way to quickly shut down a model is an important safety measure. This move by OpenAI could set a standard for emergency controls in AI research and deployment.
Key points
- OpenAI has deployed a 'kill switch' for its research AI models.
- The system, called Control Room, can halt model outputs in about 35 seconds.
- It is designed to work even if other systems fail or are attacked.
- OpenAI plans to improve the system and expand monitoring for risky behaviour.
- This highlights the need for emergency controls in AI deployments.
Sources
The Register
Reference
