Microsoft tightens AI governance as agentic risks rise

Microsoft has published its third annual Responsible AI Transparency Report, outlining changes to its governance of generative and agentic AI.

The report says Microsoft has revised its internal Responsible AI Standard as AI systems grow more complex and scrutiny of their safety, reliability and oversight increases.

Among the changes, Microsoft has reworked the standard around different parts of the AI technology stack, including models, platform services and applications. The revised framework combines baseline requirements with scenario-specific rules that can change as technical risks and regulatory demands shift.

The approach reflects a broader effort to adapt governance to AI systems that can retain memory, use tools, access data and act on behalf of users. Oversight increasingly needs to cover interactions between models, agents, applications, tools, data and people, rather than focusing only on a single model or software product.

Microsoft's controls now place greater emphasis on agent identities, tool permissions and the monitoring of actions taken by AI systems. Some of its toughest risk management measures are being applied to systems with significant cyber-related uses.

Training and tools

Over the past year, Microsoft said it provided responsible AI training to nearly 20,000 engineers, policymakers and customers. The training covered issues including threat modelling for agentic AI and defences against prompt injection.

Alongside the governance changes, the report outlines several internal and external tools intended to help developers test and monitor AI systems. These include an AI Red Teaming Agent to identify and assess risks, agent evaluators to measure quality and safety in agentic applications, and a system called RAMPART that converts red-team findings into repeatable tests.

Microsoft also described tools designed to increase visibility into agentic systems. ASSERT and Agent Control Specification are intended to let developers test agents against policies, place runtime controls at key points in workflows and monitor behaviour after deployment.

Those measures point to a shift from one-off assessment before launch to more continuous oversight after release. As AI systems become more dynamic, organisations need ways to see what they are doing and intervene when necessary, the report says.

Microsoft also said it is certified against ISO 42001 across products including Microsoft 365 Copilot, Foundry and GitHub Copilot. It has also simplified and strengthened the internal processes that support that certification.

Research links

Singapore features in Microsoft's account of its wider research partnerships on AI safety. The company said it has worked with AI Safety and Security Institutes and related bodies in Singapore, Australia, the UK and the US to strengthen AI evaluation and related technical practice.

That work spans multilingual AI performance, human-AI interaction risks and methods for testing safety. Microsoft said the partnerships are part of a wider push for shared expectations on how AI systems should be evaluated, monitored and governed across borders and sectors.

It also said it launched an External Red Team Alliance with 18 universities across six continents to improve understanding of priority risks. In parallel, Microsoft highlighted its work with industry and standards groups including the Frontier Model Forum, OpenTelemetry, the Appia Foundation, the OECD-led Hiroshima AI Process Reporting Framework and MLCommons.

One aim of that collaboration is to develop more common measures for AI risk and reliability. Microsoft said it is helping expand MLCommons' AILuminate into a broader suite of benchmarks covering areas such as jailbreak resilience, multilingual performance and psychosocial risk in conversational AI.

Industry pressure

The report lands as large technology companies face mounting pressure from governments, customers and civil society groups to show how they assess and manage the risks linked to fast-moving AI products. Agentic AI, which can perform tasks with greater autonomy, has become a particular focus because of concerns over security, misuse and loss of control.

Microsoft framed the latest changes as part of a move towards governance more closely tied to engineering work and the full lifecycle of AI systems. Responsible AI practice cannot remain static as models, products and real-world uses continue to change, it said.

"Our experience over the past year has reinforced that responsible AI cannot be static. It has to be embedded in development processes, supported by practical tools, and continually informed by what we learn," Microsoft said.