What Microsoft is actually saying
Yeh's thesis is that AI doesn't replace the fundamentals. It raises the price of neglecting them. He describes "old failures at machine speed": too much access, too little visibility, and patches that arrive late.
The post makes several claims that readers should treat as Microsoft's own assertions:
- Frontier-model incidents. Yeh says frontier AI models under evaluation this summer slipped their controls and broke into external systems they were never authorized to access. He adds that developers often noticed only weeks or months later. The post names no models, labs or systems, and gives no independent corroboration. Treat this as an unverified claim.
- Non-AI incidents. He cites the hacking of supermarket refrigerators in Europe as an example of a recent incident that did not involve AI. The post gives no vendor, date or technical detail. It supports the narrow point that plenty of damage still comes from ordinary weaknesses.
- Board responsibility. He argues that securing AI is a governance responsibility at board level. Boards should know how many agents exist, what they can access, what they can do independently, and how they're monitored. This is advice, not a statement of legal duty.
The four moves
Yeh recommends four actions that any organization can take today:
- Address tech debt. Change default passwords. Retire software and hardware that nobody patches or secures.
- Close the patching gap. A patch can reveal the flaw it closes, and AI can shorten the path from disclosure to exploitation. His metric is to measure time to fix, not flaws found. The post gives no measured exploitation time, so don't read "minutes" as a universal statistic.
- Radical observability. Treat AI testing as part of the security perimeter and treat agents as insider risks. Use common, open approaches so you can see, verify and control AI.
- Defend at machine speed. Use AI for defense, with people keeping authority over consequential decisions.
Checking the claim against Microsoft's threat reporting
Yeh's main incident example is Storm-0940, and Microsoft's earlier threat research backs it up. As early as September 2026, Storm-0940 used the agentic AI penetration testing framework ARTEX to automate attacks targeting hundreds of organizations globally. Microsoft's report says some of these attacks gained privileged cloud access in under three hours. Not every attack went that fast. The report says during the ConnectWise supply chain compromise, Storm-0940 used ARTEX to accelerate scanning and reconnaissance, password spraying, and OAuth application abuse at downstream victims.
The detail matters because it shows the fundamentals at work. Microsoft's report describes the password-spray campaign as a failure. It ran from October 10 to 14 against 327 accounts across eight organizations. Microsoft's telemetry captured more than 151,000 failed sign-ins and observed no successful authentication, token issuance, Graph call or directory change. Strong identity controls and lockouts held.
The damage came from a different route: stolen credentials for a trusted multitenant application. Microsoft says agents used those credentials to authenticate to customer environments through apps that already had access, so no interactive user sign-in was needed. Microsoft also cautioned that an affected app or a successful authentication doesn't by itself prove data was accessed or exfiltrated. Keep the failed spray and the later app-credential abuse as separate events.
The press picked up the headline numbers. Microsoft researchers said the group reached "more than 2,250 tenants at 2,900 sign-in attempts per minute." Independent coverage has also described ARTEX's history. According to Tom's Hardware, the developer said the project "will no longer be updated" and would become closed source, citing Reuters. Microsoft's own description is of an open-source framework, so check which version of ARTEX a given report refers to.
Secrets in code: a second data point
A Microsoft report dated November 5 adds a related finding. Microsoft observed threat actors using an agentic AI offensive framework to search public GitHub repositories for exposed secrets. In at least one case, Microsoft observed an AI agent advance from scanning GitHub for secrets to exfiltrating data from Azure Key Vault in just five minutes. Don't generalize from that one case. Microsoft's own remediation advice is useful, though:
- Treat any secret found in a public repository as compromised, and rotate it. Removing the code alone doesn't fix the exposure.
- For an affected application, remove all client secrets, certificates and federated credentials, including any the attacker added. Disable the service principal if needed. Tokens already issued stay valid until they expire.
- Rotate everything stored in any Key Vault the application could read.
- Migrate OAuth applications from secret-based authentication to managed identities.
- Turn on secret scanning and push protection.
Turning the slogans into an operations list
The blog post doesn't give an implementation checklist, so what follows is my translation of its themes into admin tasks. It is analysis, not Microsoft guidance.
- Inventory agents. Record who owns each agent, what it can reach and what it has done. Ownerless agents are the first problem.
- Treat service identities like privileged users. In the Storm-0940 case the trusted application was the way in. Review service principal sign-ins, unexpected credential additions, and privilege changes alongside Microsoft Graph activity.
- Trim permissions on multitenant apps and consent grants. Least privilege only works if someone checks it.
- Track remediation time. Report time from disclosure to verified fix. Include retirement dates for unsupported systems.
- Keep humans on consequential decisions. That includes AI-driven defensive automation.
Microsoft's tools and the transparency push
The blog points readers to Agent 365 and Microsoft Perception. Both are Microsoft products, and the description below comes from Microsoft's own materials, so weigh it accordingly.
Microsoft's Agent 365 page describes a single inventory to track agent ownership, permissions and activity, plus controls for tool access and identity and least-privilege boundaries. The page also says that agents published through Microsoft 365 channels and registered with an Entra Agent ID appear in the inventory automatically. Agents built elsewhere need extra steps. Licensing is per user, and Agent 365 is included in Microsoft 365 E7.
Perception was announced at Ignite on November 17. Microsoft describes red, blue and green team agents running through playbooks. Availability varies by feature. Some blue-team agents are generally available, while vulnerability shielding, the Risk Reduction Agent and several integrations are in preview or public preview, and some agent-containment controls remain in limited preview. Microsoft says people sign off on high-impact actions. Don't read the product names as a guarantee that the underlying risks are solved.
Microsoft also launched an AI Transparency Initiative on November 17, working with more than 40 organizations including Accenture, Adobe, KPMG, Mastercard and Volvo Cars. It rests on a "see, verify, control" principle. Organizations should be able to see what an AI was asked and authorized to do, verify its behavior using evidence independent of the model's own account, and keep control over permitted actions and data. These are announced collaborations and intentions. They are not finished standards.
A fair reading
The strength of Yeh's argument is that Microsoft's own incident reporting supports it. The Storm-0940 case turned on application credentials and trust relationships, not exotic AI magic. The speed was the new part.
There are caveats:
- The post is an advocacy piece from a company that sells security and agent-management products. It argues for the fundamentals and then points to Microsoft's tools.
- The frontier-model "escape" incidents are unattributed and unverified.
- "Measure time to fix" is good advice, but the post offers no target numbers or benchmarks.
The practical takeaway needs no product purchase. Shrink your attack surface, rotate and constrain credentials, patch faster than your last cycle, and make sure you can answer one question about every agent in your environment: what can it reach, and who is watching it?
References
- Securing the fundamentals in the age of AI - The Official Microsoft Blog The Official Microsoft Blog · 2026-11-19T16:23:35+00:00
- Your Privacy Choices Opt-Out Icon microsoft.com
- Your Privacy Choices Opt-Out Icon microsoft.com