Anthropic has been nominated for the 2026 World AI Awards in the Agent Governance & Policy Control category, recognising its work on the controls, safeguards and governance systems surrounding increasingly autonomous AI agents.
The category tackles a problem that has become harder as AI systems have gained the ability to do more than generate an answer. Products such as Claude Code and Claude Cowork can work across files, execute code and carry out multi-step tasks. Giving an AI agent that kind of reach creates an obvious second question: what should it actually be allowed to do?
That is where Anthropic’s work becomes particularly relevant.
The World AI Awards recognises organisations, individuals, products and technologies contributing to the development and practical application of artificial intelligence across industries.
For Anthropic, agent governance increasingly means building technical boundaries around Claude rather than relying on users to approve every action.
Claude’s growing autonomy creates a new control problem
Claude Code shows the tension clearly.
By default, Claude Code uses permissions to control potentially consequential actions. Anthropic has also developed sandboxing that restricts what the agent can access through filesystem and network boundaries.
There is a practical reason for that architecture. Anthropic reported that sandboxing reduced permission prompts by 84% in its internal use. Fewer prompts matter because constantly asking a person to approve actions does not automatically produce better oversight.
Anthropic later reported that users approved roughly 93% of Claude Code permission prompts. That creates the risk of approval fatigue: the human technically remains in control, but repeated requests can turn approval into a reflex.
The company’s 2026 Claude Code auto mode takes another approach. It uses classifiers to automate some permission decisions while retaining restrictions around actions considered more dangerous.
None of this makes autonomous agents risk-free. Anthropic itself acknowledges that probabilistic defences can miss things. The significance is more practical: governance is being pushed into the architecture of the agent instead of being treated solely as a policy document.
Sandboxing puts a boundary around what an agent can touch
Anthropic has increasingly described agent security in terms of controlling an AI system’s potential “blast radius.”
The principle is straightforward. Rather than trying to supervise every individual decision an agent makes, developers can restrict the environment in which it is capable of acting.
Claude Code sandboxing uses operating-system-level controls to create filesystem and network isolation. Files can be limited to permitted directories, while network restrictions can constrain which external services the agent reaches.
Anthropic has extended this containment philosophy across Claude products as agents have gained broader access to computers, development environments and business workflows.
That distinction matters for agent governance. A model can be highly capable while the surrounding system still determines which files it can modify, which services it can contact and how far an unexpected action can travel.
Enterprise administrators get their own policy layer
Agent governance becomes more complicated inside a company, where a single AI system may interact with proprietary code, internal documents, connected services and regulated information.
Anthropic has built administrative controls around Claude for that environment.
Claude Enterprise supports role-based permissions, audit logs, SSO, SCIM and custom data-retention controls. Administrators can therefore govern access at an organisational level rather than leaving every configuration decision to individual users.
Claude Code adds another layer through managed policy settings. Organisations can enforce settings across users, including tool permissions, file-access restrictions and MCP server configurations.
Anthropic has also introduced a Compliance API, giving enterprise customers programmatic access to Claude usage information and customer content for compliance workflows. The company says organisations can use it to integrate Claude activity into existing monitoring systems, flag potential issues and support automated policy enforcement.
That is a different kind of AI governance from model alignment. It is operational governance: deciding who gets access, what an agent can connect to, what administrators can inspect and which policies follow the agent across an organisation.
Anthropic is building governance above the product layer too
Anthropic’s approach extends beyond Claude’s user-facing controls.
Its Responsible Scaling Policy, first introduced in 2023 and repeatedly updated since, provides a framework for assessing potentially severe risks as increasingly capable models are developed and deployed.
Version 3.0, released in February 2026, introduced Frontier Safety Roadmaps and Risk Reports covering Anthropic’s deployed models. The policy has continued to change during 2026 as the company adjusts capability thresholds, reporting requirements and external review mechanisms.
Anthropic’s Frontier Safety Roadmap separately addresses security, safeguards, alignment and policy. Among its stated work are stronger safeguards against dangerous model use, systematic alignment assessments and research into security measures for increasingly capable AI systems.
The two layers should not be confused. Enterprise permissions and sandboxes govern what deployed agents can access and do. Anthropic’s Responsible Scaling Policy deals with broader questions about increasingly capable frontier models and the safeguards the company believes should accompany them.
Together, though, they point toward the same emerging challenge: AI governance increasingly has to operate at several levels at once.
Graham Cooke, President of the World AI Awards, said:
“Anthropic’s nomination in Agent Governance & Policy Control highlights one of the defining challenges emerging from the shift toward agentic AI. As AI systems gain the ability to use tools, work across applications and take more actions on behalf of people, organisations need practical ways to determine where that autonomy begins and where it stops.
“Anthropic’s work across permissions, sandboxing, enterprise policy controls and broader model governance draws attention to the importance of building oversight into the systems surrounding AI agents, rather than treating governance as an afterthought. We congratulate Anthropic on its 2026 World AI Awards nomination and look forward to following how these approaches develop as agents become more capable.”
Anthropic joins organisations, researchers, entrepreneurs and technology developers being recognised through the 2026 World AI Awards.
The programme recognises organisations, individuals and technologies contributing to the development and application of artificial intelligence across industries.
Anthropic’s nomination in Agent Governance & Policy Control lands at an interesting moment. AI agents are becoming useful precisely because they can act with less supervision. Yet the more access they receive, the more consequential mistakes, misuse or compromised instructions can become.
There is no single control that resolves that tension. Anthropic’s work instead points toward layers: model safeguards, permissions, sandboxes, access controls, monitoring and organisation-wide policies.
The agent can do more. The harder engineering problem may be making sure it cannot do everything.
Learn more about Anthropic and Claude at https://www.anthropic.com/.
Discover the World AI Awards 2026, explore the nominees and learn more about the awards at https://www.worldawards.ai/.
