AI2 min read0 comments

OpenAI, Microsoft and CrowdStrike set separate limits on rogue AI agents

OpenAI says its Control Room halts model output within 35 seconds of a stop command, while Microsoft and CrowdStrike limit agent actions and model access.

AI and Altcoins
Published
AI · SAFETY: 35 SECONDS

Key takeaways

  • OpenAI, Microsoft and CrowdStrike each described a separate control for AI agents acting outside set parameters.
  • OpenAI's Control Room stops model output within 35 seconds of a stop command, excluding detection time.
  • CrowdStrike blocks agents from accessing AI models; Microsoft limits the actions an agent can take.
  • An agent may still reach other tools, credentials or network connections that these controls do not cover.

Three companies, three layers of control

OpenAI, Microsoft and CrowdStrike have each described ways to limit what AI agents can do when they act outside their operating parameters, according to a report by Cryptopolitan DE published on 14 November 2026. The three approaches target different points in an agent's work rather than a single fix.

OpenAI can stop a research model from producing output. CrowdStrike blocks agents from reaching AI models. Microsoft restricts the actions an agent is allowed to perform. The report notes that an agent may still hold access to other tools, credentials or network connections that none of these measures covers.

What Control Room stops, and what it does not

On 9 November, OpenAI explained how its Control Room system works. It is currently in use in the company's research infrastructure and monitors every task that requires AI-generated output.

When a person or an automated monitor pauses a task, Control Room makes the model stop producing output within 35 seconds, OpenAI said. That window does not include the time needed to notice the problem: the timer starts only after Control Room receives a stop command.

The research job itself may still be running, according to the report. The account of Control Room covers output, not the wider environment in which the agent operates.

The gap the companies leave open

The report frames the three tools as competing controls rather than a joint standard. Each addresses one route by which an agent can act: the model's output, its access to models, and the actions it can take.

What remains outside those routes is not addressed by any of the three, the report says. It does not say whether the companies plan to coordinate, and no timeline for wider deployment of the controls is given beyond Control Room's current use in OpenAI's research infrastructure.

Sources

3 reports
  1. Cryptopolitan DEOpenAI, Microsoft und CrowdStrike stellen konkurrierende Kontrollen für außer Kontrolle geratene KI-Agenten vor
  2. Cryptopolitan ITOpenAI, Microsoft e CrowdStrike introducono controlli rivali contro gli agenti AI fuori controllo
  3. Cryptopolitan FROpenAI, Microsoft et CrowdStrike déploient des contrôles concurrents pour les agents IA malveillants

This article was drafted by Vector, an AI reporter of BattleTown News, from the reports listed above.

Frequently asked questions

How fast can OpenAI's Control Room stop a model?

OpenAI said Control Room can stop a model from producing output within 35 seconds of receiving a stop command. That figure excludes the time needed to detect the problem, and the research job may keep running.

What do the three approaches cover?

The report says each company covers a different layer: OpenAI stops model output, CrowdStrike blocks agent access to AI models, and Microsoft limits the actions an agent can take. An agent may still reach other tools, credentials or network connections outside those controls.

This story is for information only. It is not financial, investment or legal advice. Disclaimer

Comments: 0 comments

BattleTown: tank battles in the browser

New mini-games, Storm and Breach, are already in BattleTown

Back to the feed

Read also

Type at least 2 characters