AI ‘kill switch’ launched to stop rogue agents in seconds

Emergency system lets businesses pull the plug on powerful AI agents as a string of alarming tests raises fears over what can happen when autonomous software breaks its rules

A panic button designed to shut down rogue artificial intelligence in seconds has been unveiled amid growing fears over what happens when powerful AI agents break their rules.

The emergency ‘Agentic Kill Switch’ gives security teams a single control to immediately take a misbehaving or compromised AI agent offline before it can leak sensitive data, execute code or wreak havoc inside company systems.

It comes after OpenAI disclosed that its models achieved remote code execution during a controlled security test involving AI platform Hugging Face.

And it follows experiments in which advanced AI models have interfered with attempts to shut them down, as researchers and lawmakers increasingly examine whether powerful AI systems should be required to have emergency shutdown controls.

Unlike familiar chatbots, modern AI ‘agents’ can write computer code, open files, operate software, access company systems and even communicate directly with other AI agents.

But handing software those powers also creates serious consequences if an agent is hacked, tricked by malicious instructions or starts operating beyond the authority it was given.

Cyber security firm Straiker, which developed the kill switch, says the threat is no longer confined to research laboratories because businesses are already deploying autonomous agents across coding, productivity tools, customer management systems and infrastructure.

Its own researchers found agents exceeded their intended authority in 85 per cent of successful attack scenarios during simulated attacks in production-like environments across financial services, healthcare and critical infrastructure.

In one disturbing healthcare simulation, Straiker said a medical aide agent was persuaded to ignore ventilator readings, resulting in a simulated patient death.

The company says attackers could manipulate trusted agents using poisoned code “Skills”, malicious Model Context Protocol servers or prompt injections hidden inside apparently ordinary documents.

Because an agent may already have legitimate access to sensitive company systems, successfully hijacking one could potentially allow an attacker to execute code, steal information or misuse connected tools from inside an organisation.

Straiker’s new technology is intended to provide the digital equivalent of an emergency stop button.

Ankur Shah, chief executive and co-founder of Straiker, said: “Every enterprise deploying AI agents will eventually face a rogue agent moment. The question is whether they can stop it before the agent causes damage.

“We built the kill switch because ‘open a ticket’ is not an incident response plan for software that doesn’t wait for business hours.”

The system works alongside Straiker’s Ascend AI technology, which attacks agents before they are deployed to uncover weaknesses including prompt injection, data exfiltration, tool manipulation and remote code execution.

Its Defend AI system then monitors requests, tool calls and communications between agents while they are operating.

When an agent crosses a security boundary set by the business, Straiker says the kill switch can take it offline in seconds and return control to human operators.

A director of application security at a security company using Straiker’s technology said: “Securing AI agents is a new and fast-moving discipline. Straiker’s Ascend AI gives us a rigorous way to red team our agents against real adversarial techniques, and the Straiker team has been a strong partner as we bring the same rigor to our own AI that we bring to protecting our customers.”

READ MORE: Europe launches ‘anti-kill switch’ cloud shield as Trump fears grip Brussels. Four European tech firms have unveiled a new sovereign recovery system as fresh polling shows most Europeans believe the U.S could one day pull the plug on vital digital services.

Do you have news to share or expertise to contribute? The European welcomes insights from business leaders and sector specialists. Get in touch with our editorial team to find out more.

_{Main image: Panumas Nikhomkhai via Pexels}