OpenAI, Microsoft and CrowdStrike detail controls for AI agents
OpenAI, Microsoft and CrowdStrike have described separate controls intended to contain risks from AI systems and agents. Their approaches range from stopping research workloads to blocking agent actions or cutting network access.
Key takeaways · 4
- 01
OpenAI says Control Room connects research inference workloads and stops output within 35 seconds after a decision to pause or stop a job.
- 02
OpenAI requires an approved safety case for every research workload that uses agentic inference.
- 03
CrowdStrike says its kill switch can stop internal agents, block relaunches and cut enterprise network access.
- 04
Microsoft says its new capabilities can assess agent risk and block changes, require human intervention or shut an agent down.
OpenAI focuses on research workloads
OpenAI says Control Room tracks model workloads across its research infrastructure and can stop them when safety or security concerns arise.[1] It connects every research inference workload, including agentic sampling for training and evaluation, data processing and other workloads.[1] OpenAI says the system halts model output within 35 seconds of a decision to stop or pause a job.[1] It is already deployed across the research infrastructure, where OpenAI says the need is greatest.[1] The company says the effort responds to the importance of detecting issues and stopping models as capabilities and risks grow.[1]
Controls include checks and safeguards
OpenAI says inference engines check with Control Room before fulfilling a request, and the service denies requests for jobs that have been stopped.[1] The company also says Control Room is designed to keep working if other systems fail and to stop model output if Control Room itself goes down.[1] Separately, OpenAI requires an approved safety case for each research workload using agentic inference.[1] OpenAI says its external model-serving stack is more centralized and has the operational capability to take down deployments quickly if needed.[1]
CrowdStrike offers an agent kill switch
CrowdStrike announced an AI kill switch and said it was available to Falcon Guardian customers.[2] The company says the switch can stop internal agents, block them from relaunching and cut their access to an enterprise network.[3] CrowdStrike also said it was introducing security innovations to strengthen AI security architecture, expand cloud coverage and enable agent insider-threat detection.[2] It framed this work as spanning endpoint to cloud.[2] These announced functions differ from OpenAI's stated focus on tracking and stopping research inference workloads.[1][3]
Microsoft describes agent-level options
Microsoft said new capabilities in Microsoft 365 E7 and Project Perception would help organizations contain agent risk and adapt protection continuously.[4] The announced capabilities include assessing each agent's risk and, depending on the situation, blocking it from making changes, requiring human intervention or shutting it down.[5] The announcement therefore describes several possible responses rather than one universal shutdown action.[5] Teams evaluating these announcements should distinguish controls for research infrastructure, internal agents and organization-managed agents.[1][2][4]
Teams assessing agent controls should compare what each mechanism can stop, where it applies and what happens when a control fails. The announcements offer distinct approaches, so professionals should verify that a proposed control matches the workloads and access paths they need to govern.
Why it matters
AI training for your team
kju.ai is built for teams: daily sessions, team analytics, and rollout support for your organization.
How this developed
15 November 2026
OpenAI, Microsoft and CrowdStrike detail controls for AI agents
Sources
- Control Room: A central system to track and stop model workloads across our research infrastructure | OpenAIopenai.com
- CrowdStrike Advances AI Security and Safety from Endpoint to Cloudcrowdstrike.com
- CrowdStrike launches AI kill switch to control internal agentscnbc.com
- Agents are your newest insiders. Security must move at their speed. | Microsoft Security Blogmicrosoft.com
- Microsoft unveils tools to monitor and corral rogue AI agents – GeekWiregeekwire.com
