Microsoft stated that through the new AI Transparency Initiative, the company is collaborating with over 40 organizations, including Accenture, Adobe, KPMG, Mastercard, and Volvo Cars, to establish a common set of AI transparency, security, and governance practices among different models and providers. Microsoft will also enhance the visibility of how AI models and proxies interact with corporate data in its products and services.
Redmond, Washington, November 17th / PRNewswire / -- Organizations can observe most of the activities surrounding the AI model, but it is often difficult to reconstruct what happens between requests and results, which makes it challenging to independently verify the behavior of the model. On Tuesday at the Microsoft Ignite conference, Satya Nadella, Microsoft's Chairman and CEO, announced the next phase of Microsoft's commitment to making AI more secure: the AI Transparency Initiative. This is an effort in collaboration with customers and partners aimed at establishing common, open methods of security and governance across models and providers. The initiative embodies a simple principle: the greater the AI autonomy, the higher the transparency must be.
"For security controls to be truly effective, every organization needs to have greater transparency regarding the models themselves," said Hayete Gallot, Executive Vice President of Microsoft Security. "With this effort, we will bring more transparency to all models. This is a key step in making containment measures smarter and more powerful, which will help us improve accountability and traceability even further."
Enterprises are connecting the AI model with data, tools, credentials, and infrastructure, enabling agents to take action without requiring manual approval at each step. Existing control measures can observe most of the activities surrounding the AI model, but they cannot see what is happening within the model, nor can they independently verify the model's explanations for its own behavior. This limits an organization's ability to understand the context behind the agents' actions. This gap in evidence poses risks to security, compliance, auditing, and accountability. Security teams need a new trust architecture to address the growing internal risks posed by increasingly powerful AI agents.
CISO Aanchal Gupta stated: "As AI proxies become more deeply integrated into business operations, the security team needs to have a clear understanding of what these proxies access, what actions they take, and how those actions are verified. Establishing a common set of AI transparency practices across industries is crucial for enhancing trust and helping organizations to innovate with confidence while maintaining security and accountability."
From Principles to Practice
The initiative is divided into three parts. First, Microsoft will expand its participation with the National Institute of Standards and Technology (NIST) and utilize its research to develop open, industry-level standards for transparency, security, and governance. Second, Microsoft is collaborating with more than 40 organizations, including Accenture, Adobe, KPMG, Mastercard, and Volvo Cars, to help establish open standards that ensure that AI systems are accountable, independently verifiable, and inherently secure regardless of the model or provider used.
Thirdly, Microsoft will incorporate this work into Microsoft AI, Microsoft Foundry, and Microsoft Copilot, enabling customers to use the resulting evidence in conjunction with their security and compliance tools, whether these tools come from Microsoft or other providers.
"Our customers want to adopt AI more quickly; they shouldn't have to choose between innovation and security," said KPMG International, Global Chief Information Security Officer of John Israel. "Trust in AI requires transparency, explainability, and responsible oversight. We are proud to collaborate with Microsoft to advance the AI transparency initiative, helping to establish open, industry-wide standards that provide organizations with the visibility and control needed for large-scale, responsible deployment of AI."
Clear accountability
Proxies that have access to data, tools, and credentials have a risk profile similar to that of privileged insiders, but they operate at machine speed and scale; therefore, independent oversight is of paramount importance.
Accountability still lies with the organization itself. When significant events occur, leaders need to be accountable to customers, employees, the board of directors, auditors, and regulatory authorities. Blaming AI is not the solution, nor is "believing in the model" a means of control.
The initiative is designed to operate across models and providers, including both closed models and open-weight models. This allows development teams to replace or update the models that power their agents without compromising the ability of the IT team and security teams to govern them securely, nor does it diminish the capability of compliance teams to audit their behavior.
Accenture The global head of cybersecurity, Harpreet Sidhu, stated: "As organizations take the next step in AI reshaping, they need a solid foundation of trust and confidence. Microsoft's open and industry-driven approach to initiatives such as AI transparency can help us go further in protecting AI models and equipping them with control and governance. Accenture is proud to collaborate with Microsoft and other leaders in this field."
See, verify, control
Microsoft believes that every organization should be able to see what the AI system has accessed and done, independently verify its behavior, and control what it is allowed to do.
- See:Understand what AI is required to do, what it is authorized to do, which models and runtimes are used, which data and tools are accessed, and what actions are taken. Continuously monitor proxy activities and risk signals, and compare them with policies and thresholds.
- Verification:Agents cannot be the sole source of truth for their own actions, nor can frontier models be trusted to govern themselves. Evidence that encompasses agent identity, intent of delegation, and actions should indicate its source (the model, the runtime surrounding it, or independent systems such as identity and strategy), in order to verify the signals reported by the models against independently observed records. Organizations should be able to use their own security and compliance systems to independently verify AI activities, without having to rely entirely on the model providers' explanations for their behavior.
- Control:Maintain control over what AI is allowed to do. The purpose, owner, data, and delegated permissions of each workflow should be clearly documented to assist the organization in implementing policies before significant actions are taken, and to reconstruct what occurred during investigations, containment, recovery, and audits. Even if the models or providers change, this documentation should remain in the organization's possession. Regardless of any changes to the models or providers, the organization should retain control over the data that its agents have access to and the methods by which they can access it.
How to Participate
Organizations interested in helping to shape this initiative or tracking its progress can visit aka.ms / learnmoreAItransparency to learn more and register.
Microsoft states that it welcomes attention to the news on Microsoft Ignite. Microsoft and industry leaders will share insights on how organizations can confidently manage AI models and proxies at the conference. For more information, please visit the Microsoft Security blog.
Note to the Editor
For more information, news, and insights from Microsoft, please visit Microsoft Source : https :// news.microsoft.com / source. The web links, phone numbers, and job titles mentioned in this article were accurate at the time of publication, but may have since changed. Reporters and analysts who need further assistance can contact Microsoft's Rapid Response Team or other appropriate contacts listed at https :// news.microsoft.com / microsoft-public-relations-contacts.









