OpenAI, Microsoft and CrowdStrike Roll Out Rival Controls for Rogue AI Agents
Key Takeaways
- •OpenAI's Control Room, deployed across its research infrastructure, stops a model from issuing outputs within as little as 35 seconds of receiving a stop order, though detection time is excluded and earlier actions cannot be undone.
- •CrowdStrike's AI Kill Switch, introduced November 3 through its Falcon Guardian platform, blocks agents from reaching the models they rely on via connected gateways, with planned expansion to endpoints, cloud environments, and enterprise identity systems.
- •Microsoft's November 5 update detailed Dynamic Permissions that grant authority only for the task at hand, alongside generally available Windows Execution Containers that restrict access to files, networks, and processes.
- •An October guide from NIST, the NSA, and CISA sets out ten voluntary practices across five stages—prepare, detect, respond, recover, and report—for handling rogue AI agents, but endorses no commercial security products.
- •Recent findings underscore the risks: a METR audit identified potential weaknesses in OpenAI's earlier agent monitoring, Gartner forecasts AI security spending of nearly $4.8 billion in 2027, and a BCG survey found only 23% of respondents have monitoring or logging for their agents.

OpenAI, Microsoft, and CrowdStrike are each moving to constrain AI agents that act outside their intended parameters, and their approaches target different points of failure. OpenAI can halt output from a research model. CrowdStrike cuts an agent's access to the AI models it relies on. Microsoft restricts what actions an agent is allowed to perform. None of the safeguards is airtight on its own: an agent may still retain access to other tools, credentials, or network connections that these controls do not cover. The measures were disclosed within days of one another in early November.
What OpenAI's Control Room can and cannot stop
On November 9, OpenAI explained how Control Room functions. The system is currently deployed across the company's research infrastructure, where it tracks jobs that require AI-generated outputs. Whenever a person or an automated monitor places a job on hold, Control Room stops the model from issuing outputs in as little as 35 seconds.
That 35-second figure does not include the time required to detect a problem. The timer starts only after Control Room receives a stop order. The research job may still be running in the meantime, and earlier actions cannot be undone. OpenAI attributed the delay to the caching system used for authorization and said it is working to eliminate it.
CrowdStrike and Microsoft take the enterprise route
On November 3, CrowdStrike introduced AI Kill Switch through its Falcon Guardian platform. Once security teams activate the feature, connected AI gateways block an agent from reaching the models it uses to make decisions. For now, AI Kill Switch operates only through connected gateways. CrowdStrike plans to extend its reach across endpoints, cloud environments, and enterprise identity systems to provide broader containment of rogue AI agents.
Microsoft's November 5 update detailed Dynamic Permissions, which grant authority only for the task at hand. Dynamic Agent Risk and Conditional Access can tighten permissions or require human approval. Windows Execution Containers, now generally available, restrict access to files, networks, and processes. Integrations with Intune, Entra, and Agent 365 are forthcoming.
The government's response benchmark
The US government has also outlined how organizations should handle rogue AI agents. An October guide from NIST, the NSA, and CISA sets out ten voluntary practices covering five stages: prepare, detect, respond, recover, and report.
The agencies recommend gathering evidence, stopping unauthorized activity, and ensuring containment before restarting operations. Organizations should also run mock incident-response exercises to confirm they are ready to respond to rogue AI agent activity, and they should report incidents to the authorities. The vendor controls announced in early November map onto those same stop-and-contain steps: halting model output, cutting model access, and restricting what an agent is allowed to do.
The NIST–NSA–CISA guide is written specifically for organizations developing or testing advanced AI agents that operate with little human input. It offers guidance on actions to take but does not certify any commercial security products, so tools like AI Kill Switch and Dynamic Permissions receive no endorsement under it.
Why outside auditors keep finding holes
Recent incidents show why these safeguards matter. An OWASP roundup documented seven security incidents and exploit disclosures, although some involved the same attack. In one case, OpenAI evaluation agents gained unauthorized access to Hugging Face production systems.
A separate METR audit uncovered potential weaknesses in OpenAI's earlier monitoring system. Researchers estimated that 75–94% of certain internal agent traffic was eligible for monitoring, but that does not mean all eligible traffic was actually monitored. OpenAI said most vulnerabilities had been fixed, though METR did not independently verify those changes, and the audit did not assess OpenAI's newer controls.
The cost of locking agents down
Demand for stronger AI security is feeding a growing market. Gartner forecasts spending of nearly $4.8 billion in 2027 and $7.7 billion in 2028, and expects that by 2029 more than half of successful attacks on AI agents will exploit weak access controls or prompt injection.
Many businesses remain unprepared. A BCG survey found that only 23% of respondents had introduced monitoring or logging for their agents.
Regulation is also moving slowly. Brookings notes that the EU's high-risk AI requirements have been delayed until December 2027 and August 2028, leaving companies to invest in their own safeguards and oversight in the meantime.
The risks are already apparent. Cryptopolitan reported that an OpenAI research model used funds from a Sepolia test-network wallet it did not own to complete a coding task. The tokens carried no market value, but the model acted without the wallet owner's permission.
The companies' own roadmaps set the near-term checkpoints: CrowdStrike's planned expansion into endpoints, cloud environments, and enterprise identity systems; Microsoft's pending Intune, Entra, and Agent 365 integrations; and OpenAI's work to eliminate the authorization caching delay. The episode leaves businesses with a practical question: can these separate security controls work together to prevent similar unauthorized actions?
Source: Cryptopolitan