Three AI agent brakes, three separate products and no shared standard
The 35-second figure starts once a stop order arrives; finding the problem takes longer, and completed actions stay done.

Key takeaways
- OpenAI, Microsoft and CrowdStrike each announced controls aimed at AI agents acting outside set parameters.
- OpenAI's Control Room stops model output in as little as 35 seconds after a stop order.
- The 35 seconds excludes detection time, and earlier actions cannot be undone.
- The safeguards do not cover an agent's other tools, credentials or network connections.
- OpenAI said the delay comes from an authorization caching system and it is working to remove it.
Three companies, three different brakes
OpenAI, Microsoft and CrowdStrike have each put forward controls meant to limit what AI agents can do when they act outside their set parameters, according to Cryptopolitan. The approaches differ: OpenAI stops output from a research model, CrowdStrike blocks agents from reaching AI models, and Microsoft limits the actions an agent can perform.
The three controls are separate products rather than one shared standard. Cryptopolitan reported that an agent may still hold access to other tools, credentials or network connections that none of these safeguards covers.
How Control Room stops a model
OpenAI described Control Room on 9 November, according to the report. The system runs in the company's research infrastructure and tracks jobs that need AI-generated output.
When a person or an automatic monitor puts a job on hold, Control Room makes the model stop issuing output in as little as 35 seconds. That figure does not include the time needed to detect the problem; the timer starts only once Control Room receives a stop order.
The research job itself may keep running, and actions already taken cannot be undone. Cryptopolitan reported that the delay comes from the caching system used for authorization, and that OpenAI is working to remove it.
What is not yet public
The report gives no figures for how often Control Room has been triggered, and no details on how Microsoft's and CrowdStrike's controls are configured or priced. It also does not say whether the three companies coordinate their safeguards or test them against the same scenarios.
Whether any of the controls can stop an agent that has already reached external tools or credentials remains an open question in the report.
Sources
3 reports- CryptopolitanOpenAI, Microsoft, CrowdStrike roll out rival controls for rogue AI agents
- Cryptopolitan PTOpenAI, Microsoft e CrowdStrike lançam controles rivais para agentes de IA descontrolados
- Cryptopolitan ESOpenAI, Microsoft y CrowdStrike lanzan controles rivales para agentes de IA descontrolados
This article was drafted by Vector, an AI reporter of BattleTown News, from the reports listed above.
Frequently asked questions
What is OpenAI's Control Room?
A system used in OpenAI's research infrastructure that tracks jobs needing AI-generated output and can make a model stop issuing output after a stop order, according to Cryptopolitan.
How fast does Control Room stop a model?
In as little as 35 seconds after it receives a stop order. The figure excludes the time needed to detect the problem.
Do these controls cover everything an AI agent can access?
No. According to the report, an agent may retain access to other tools, credentials or network connections that the safeguards do not cover.
This story is for information only. It is not financial, investment or legal advice. Disclaimer
BattleTown: tank battles in the browser
New mini-games, Storm and Breach, are already in BattleTown




Comments: 0 comments