Microsoft expands AI oversight in 2026 transparency report

Microsoft has published its 2026 Responsible AI Transparency Report, the company's third annual account of its work on AI governance and oversight.

The report outlines how Microsoft has revised its internal standards, expanded technical tools for testing and monitoring AI systems, and broadened its work with outside groups on shared benchmarks and standards.

At the centre of the update is a reworked Responsible AI Standard, redesigned to reflect changes in the AI technology stack and in how AI products are developed and deployed. The standard now separates requirements for models, platform services and applications, as well as by Microsoft's role in building or deploying them.

That structure combines baseline rules with more specific requirements for particular scenarios. According to Microsoft, the approach is meant to let its controls evolve as technical risks, use cases and regulatory demands change.

Microsoft also pointed to closer links between governance and engineering. Its risk management practices now focus more on systems that can retain memory, use tools, access data and take actions on behalf of users, particularly as agentic AI becomes more common.

For those systems, governance must address interactions among models, agents, applications, tools, data and people, rather than only reviewing the behaviour of a single model or application. That has led Microsoft to place greater emphasis on controls such as agent identities, tool permissions and action monitoring.

According to the company, thousands of engineers and Product Managers have also received training, including on agentic AI threat modelling and prompt injection defences. It described this as part of a shift towards continuous, lifecycle-based governance rather than a one-off assessment before deployment.

Tools expanded

Alongside the policy changes, Microsoft detailed a broader set of tools for evaluating and controlling AI systems across their lifecycle. These include a new AI Red Teaming Agent, agent evaluators for measuring the quality, safety and performance of agentic applications, and RAMPART, which turns red team findings into repeatable tests.

Microsoft also highlighted ASSERT and Agent Control Specification, designed to help developers test agents against internal policies, place controls within an agent workflow, and monitor behaviour during operation.

Microsoft presented these additions as a response to a more complex operating environment for developers and organisations using AI. As systems become more dynamic, oversight must move beyond pre-deployment review towards continuous testing, visibility and intervention.

Another area of emphasis is formal certification. Microsoft said it is certified against ISO 42001 across products including Microsoft 365 Copilot, Foundry and GitHub Copilot, and that it has streamlined the internal processes supporting that certification over the past year.

External work

Beyond its own products and internal controls, Microsoft said it has increased its work with research bodies, standards groups and universities. It said it had advanced work with the US Centre for AI Standards and Innovation and AI Safety and Security Institutes in Australia, Singapore and the UK on AI evaluation science and practice.

Microsoft also said it launched an External Red Team Alliance with 18 universities across six continents to broaden research into priority risks. That effort sits alongside work through groups including the Frontier Model Forum, OpenTelemetry and the Appia Foundation on cyber benchmarks, observability and AI assurance.

Transparency reporting and measurement were also prominent in the report. Microsoft said it is contributing to efforts to make transparency reporting more interoperable across organisations and jurisdictions, including work through an OECD-led informal task force that developed version 2.0 of the Hiroshima AI Process Reporting Framework.

Microsoft argued that consistent measurement remains a major gap in the sector, saying progress cannot be properly assessed if each organisation uses different ways to measure AI risks.

For that reason, Microsoft said it is working with MLCommons to expand AILuminate into a broader suite of reliability benchmarks. Areas cited include jailbreak resilience, multilingual performance and psychosocial risk in conversational AI.

The report also reflects broader pressure on large technology companies to explain how they are building and governing AI systems as regulators, customers and researchers demand more detail on testing, controls and accountability. Microsoft said trust in whether systems operate reliably and securely is becoming a basic condition for wider adoption.

It added that its responsible AI programme has been in place for nearly a decade and that its recent work is shaped by three areas of investment: adaptive governance and technical risk management, practical tools, and shared practices with outside partners.

Microsoft said the spread of agentic AI has made that work more urgent because risks can change as systems interact with users, environments and other systems. It added that its governance will need to keep adapting in line with those changes and with lessons learned from testing and deployment.