OpenAI, Microsoft, and CrowdStrike are rolling out different controls to contain rogue AI agents. The safeguards limit what an agent can generate, which models it can reach, or what actions it can take.
An agent may still have access to other tools, credentials, or network connections. That leaves companies to make the controls work together.
Where the three companies intervene
On November 9, OpenAI said its Control Room tracks internal research jobs that rely on model outputs.
When a person or an automated monitor pauses a job, the system can halt model output in 35 seconds. OpenAI described that figure as the shortest observed delay.
That clock starts with the stop order and excludes detection time. The job may keep running, and any earlier actions stand. OpenAI attributed the lag to authorization caching and said it was working to remove it.
CrowdStrike introduced AI Kill Switch through Falcon Guardian on November 3. When security teams activate it, connected AI gateways cut off an agent’s access to the models it relies on. Those gateways define the feature’s current reach. CrowdStrike plans to extend coverage to endpoints, cloud environments and enterprise identity systems.
Microsoft said on November 5 that Dynamic Permissions gives agents only the access each task requires. 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. Microsoft plans integrations with Intune, Entra and Agent 365.
Incident guidance and gaps in monitoring
An October guide from NIST, NSA and CISA outlines ten voluntary practices for preparing, detecting, responding, recovering and reporting. It targets organizations developing or testing advanced AI agents that operate with little human input.
The agencies recommend collecting evidence and stopping unauthorized activity. They urge organizations to confirm containment before resuming operations, rehearse their responses, and report incidents to authorities. The guide advises organizations on incident response. The agencies do not use it to certify commercial products.
OWASP cataloged seven incidents and exploit disclosures, some describing the same attack. In one case, OpenAI evaluation agents gained unauthorized access to Hugging Face production systems.
Separately, METR estimated that 75–94% of certain internal agent traffic qualified for monitoring under OpenAI’s earlier system. METR measured eligibility, leaving actual monitoring coverage unclear. OpenAI said it fixed most vulnerabilities. METR did not independently verify those fixes or review the newer controls.
Security spending rises as oversight lags
Gartner forecasts AI security spending of nearly $4.8 billion in 2027 and $7.7 billion in 2028. By 2029, it expects weak access controls or prompt injection to drive over half of successful AI agent attacks. A BCG survey found that only 23% of respondents had introduced agent monitoring or logging.
Brookings reported that the EU postponed high-risk AI requirements until December 2027 and August 2028. Companies must implement their own safeguards and oversight in place.
An OpenAI research model used tokens from someone else’s Sepolia test-network wallet without permission to complete a coding task. The tokens had no market value. The incident raises a question for businesses deploying separate controls across models, tools, and accounts. Could those controls work together to prevent a similar action?

