On May 18, 2026, a coalition led by Access Now — joined by Amnesty International, the Electronic Frontier Foundation, 7amleh and Fight for the Future — sent a follow-up letter pressing Microsoft to publish the factual findings of the external review it commissioned in August 2025 into the Israeli military's use of Azure cloud and AI services. The letter notes that Microsoft has confirmed the review is complete, yet has disclosed neither the report itself nor the scope of its investigation, the entities it covered, or which services were ultimately restricted in the company's September 2025 suspension targeting Israel's military intelligence Unit 8200.
The sequence matters because it illustrates a structural problem with how the cloud-and-AI industry currently handles defense contracts: voluntary transparency without a backstop.
The 12-month credibility gap
On May 15, 2025, Microsoft published a statement saying that after an internal review and additional fact-finding by an external firm — "interviewing dozens of employees and assessing documents" — it had found "no evidence to date" that Azure and AI technologies had been used to target or harm people in Gaza. Less than three months later, an August 6, 2025 joint investigation by The Guardian, +972 Magazine and Local Call reported that Unit 8200 had used a segregated area of Azure — with data hosted in Microsoft's Netherlands and Ireland regions — to store roughly 200 million hours of intercepted Palestinian phone calls, used to inform targeting decisions.
Microsoft then engaged Covington & Burling LLP for a formal external review and, on September 25, 2025, announced it had "ceased and disabled" a specific set of cloud storage and AI subscriptions to a unit within the Israeli Ministry of Defense. Vice chair Brad Smith said the platform had been used in a way that violated Microsoft's own terms of service, which prohibit mass surveillance of civilians.
In other words: the company's first review said there was no evidence of harm. Four months later, the company's second review concluded there had been a clear-cut terms-of-service violation serious enough to terminate live subscriptions.
That is not a small drift. It is a complete reversal — and it happened only after independent journalists reconstructed what the company's own internal processes had apparently missed.
The strongest case for Microsoft's silence
Before criticizing the company's posture, it is worth steelmanning it. Cloud providers operate under contractual confidentiality with government customers, and a defense ministry is an unusually sensitive counterparty. Releasing a forensic review in full could expose operational details — data center geography, segregation architecture, integration patterns — that adversaries could exploit. It would also create discoverable evidence in active and threatened litigation across multiple jurisdictions, including pending UN-level inquiries. Microsoft's lawyers have a legitimate professional duty to advise against publishing material that could prejudice the company's defenses or harm employees identified in the record. None of that is bad faith.
The problem is that this is exactly the rationale every cloud provider, every defense contractor, and every dual-use AI vendor will invoke. If "trust us, we reviewed it" is the entire compliance architecture for the most consequential AI systems being deployed on Earth, the architecture is too thin.
Proportionate disclosure, not blanket transparency
The right policy answer is not to nationalize the audit function or to force hyperscalers to publish every customer contract. That would chill the open commercial cloud model that has, on balance, been an enormous force for productivity, security and access. It would also collapse the legitimate confidentiality interests of millions of non-state customers who have nothing to do with armed conflict.
A more proportionate approach has three elements.
First, scoped mandatory disclosure for defense and intelligence contracts above a value threshold: not the underlying intelligence product, but the categories of services delivered, the existence and completion of human-rights due-diligence reviews, and a redacted summary of any terms-of-service enforcement actions taken. The EU AI Act's high-risk transparency obligations and the UN Guiding Principles on Business and Human Rights already point in this direction; the gap is enforcement on cloud-as-a-service, where most of the actual compute now lives.
Second, standardized terms-of-service incident reporting. Microsoft's September 2025 finding that Unit 8200 had violated the acceptable-use policy is precisely the kind of event that, in any other regulated industry — banking, telecom, medical devices — would trigger a public incident notice with a defined factual perimeter. Cloud providers should adopt a similar norm voluntarily or face it being legislated.
Third, independent technical attestation, not just law-firm fact-finding. A Covington & Burling review interviews witnesses and reads documents. It does not, on its own, audit ingestion volumes, model invocation logs, or geographic data flows. For high-risk customers, an SOC-2-style independent technical attestation — with findings shared with a designated regulator under confidentiality — would close the gap between what the company says happened and what its infrastructure actually shows.
Why this matters beyond Israel
The Azure–Unit 8200 case is not a one-off. Every major hyperscaler now sells comparable services to multiple defense and intelligence customers across the US, EU, Gulf, and Asia-Pacific. The combination of state demand for AI-enabled analytics and commercial pressure to win those contracts is not going to fade. If the industry's only disclosure mechanism for the most sensitive contracts remains a company-commissioned review whose findings the company may or may not release, the credibility of every future "we reviewed it and found nothing" statement will be lower than the last.
The coalition's May 18, 2026 letter should be read in that frame. It is not, on its face, an anti-Microsoft document. It is a stress test of whether voluntary corporate transparency can carry the regulatory weight currently placed on it. So far, the evidence is that it cannot — and proportionate, narrowly-scoped disclosure rules are now the more credible path than either blanket secrecy or blanket exposure.