Straiker launches kill switch for enterprise AI agents

Straiker has launched an Agentic Kill Switch for enterprise AI agents, designed to let security teams take a compromised or misbehaving agent offline within seconds.

The launch brings a control associated with frontier AI safety to companies already using coding agents and agent-building platforms. Support is available across enterprise coding agents and agent builder systems.

The move comes as shutdown controls for AI systems have become part of a broader policy and security debate. Straiker ties that discussion to recent concerns over model safety and the spread of autonomous agents across corporate software, data, and infrastructure environments.

Operational control

The new switch is intended for security practitioners when an AI agent appears compromised or starts acting outside its assigned role. It is launching alongside updates to Straiker's agent discovery and autonomous adversarial testing tools.

Together, those additions are meant to help companies identify which agents are running in their environments, test how those agents might be manipulated, and disable them if an incident occurs. Straiker argues that as agents gain broader access to code repositories, business applications, and internal systems, incident response needs to be faster and more direct.

That reflects a shift in how security teams view AI agents. Rather than treating them as experimental tools, many companies are beginning to treat them as operational software with access to sensitive workflows and permissions.

"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," said Ankur Shah, Chief Executive Officer and Co-Founder of Straiker.

"We built the kill switch because 'open a ticket' is not an incident response plan for software that doesn't wait for business hours," Shah said.

Attack scenarios

Findings from Straiker's STAR Labs testing helped shape the product launch. The company has been red-teaming enterprise AI agents in production-like settings across financial services, healthcare, and critical infrastructure.

In those assessments, agents exceeded their intended authority in 85% of successful attack scenarios, according to Straiker. In one simulated healthcare exercise, a medical aide agent was persuaded to ignore ventilator readings, leading to a simulated patient death.

Straiker says the risks can stem from several points in an enterprise environment, including poisoned skills, malicious Model Context Protocol servers, and prompt injections hidden in documents. Any of these could push a trusted agent into executing code, exposing data, or misusing connected tools.

For businesses deploying agents across customer systems, productivity software, supply chains, and infrastructure, the practical concern is the breadth of access these systems can gain. A single compromised agent with wide permissions could affect a large part of an organisation's technology stack.

Before and during use

Straiker says its wider platform works in two stages. Before deployment, its Ascend AI system red-teams agents to identify routes to remote code execution, data exfiltration, tool manipulation, and prompt injection.

That process is intended to show security teams how an agent can fail before it is released into live use. At runtime, its Defend AI product monitors requests, tool calls, and agent-to-agent interactions for the same attack techniques identified during testing.

When an agent crosses what Straiker describes as a defined security boundary, the kill switch can take it offline and return control to the security team. Straiker is positioning that as an operational safeguard, not just a governance measure.

One customer cited the testing element of the platform in support of the approach.

"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," said the Director of Application Security at a leading security company.

Growing market

Straiker is part of a growing group of security vendors focused on AI agents rather than conventional model monitoring alone. The sector has attracted investor interest as companies adopt tools that can write code, handle internal workflows, and interact with external systems with limited human intervention.

Straiker says it has raised USD $85 million in total funding, including a USD $64 million Series A backed by Marathon, Bain Capital Ventures, and Lightspeed. Its customer base includes Fortune 500 companies and frontier AI labs.

The market question now is whether shutdown controls become standard for enterprise AI operations, much as access controls, logging, and endpoint isolation became standard in earlier waves of enterprise security. Straiker argues that businesses using autonomous agents need a direct way to stop them when they breach their remit rather than relying on slower internal escalation.