AI is moving deeper into real services, from screen performances to legal work and desktop administration. Today’s developments show why consent, task-level permissions and measurable business results need to be designed into adoption, not added afterwards.
UK actors win proposed project-specific rights over AI replicas
Equity and Pact have agreed proposed protections giving UK screen performers consent and payment rights when digital replicas are used. The three-year deal will go to a ballot of 8,000 Equity members, with union leaders recommending approval.
Contracts would authorise specific projects rather than blanket use of a likeness. However, Equity says the agreement does not establish explicit contractual protection against using performances to train AI systems.
Our take: For businesses commissioning synthetic presenters or cloned voices, permission to create an asset should not be treated as permission for every future use. Specify the project, payment, reuse and training rights separately. This is a proposed industry agreement, not a new law covering all UK businesses.
London AI law firm Alecto raises £15 million
Alecto has raised a £15 million seed round led by Insight Partners, BusinessCloud reports. Founded in September 2025, the London consumer law firm initially focuses on residential conveyancing and plans to expand into other areas of law and launch internationally in 2027.
The company says it reached $8 million in annualised revenue within eight months of trading, while its team grew from three people in January to 85. Those figures are company claims, not independently audited results presented in the report.
Our take: The competitive change is not just lawyers buying chatbots. New firms are redesigning the service around quicker progress, clearer communication and outcomes. Established professional-services firms should measure those customer-facing results, while retaining qualified supervision and responsibility for advice.
Microsoft launches specialist models for coding and security
At Ignite, Microsoft announced new models for coding, cybersecurity and the small decisions that keep agent workflows running. MAI Code 2 leads its agentic coding family; MAI Cyber 2 Flash has 12 billion active parameters, according to Microsoft.
Microsoft-Decision-1.1 targets routing, classification, prioritisation and verification rather than sending every small task to a large general-purpose model. Performance and cost comparisons in the announcement are vendor claims.
Our take: For UK businesses, model choice should follow the task. Test a smaller specialist on repetitive decisions and measure the cost of correct, usable outputs, including retries and human checks. A cheaper model that routes work badly can cost more downstream than it saves.
Agent 365 adds business-outcome and spending controls
In a separate Ignite product announcement, Microsoft introduced Agent 365 Business Analytics to connect AI spending and adoption with business outcomes. Measures include success rate, user satisfaction and assisted hours, alongside reporting and recommendations.
Agent 365 Cost Management gives administrators spending policies, alerts and forecasting tools. These are announced capabilities, not evidence that a particular organisation has achieved savings.
Our take: Usage is not return on investment. A sensible pilot pairs spending data with work completed, quality and the cost of supervision. Ask which capabilities are included in your licence and available in your tenant before assuming an announcement changes your existing service.
Proton brings permission-controlled AI tasks to desktop
Proton has launched Lumo Desktop for Windows, macOS and Linux, with Connectors, Schedules and Skills. Connectors can access local files and tools, Proton Mail through Bridge, third-party services and local or remote MCP servers.
Users choose which actions can run automatically, which need approval and which are blocked. Proton says connected information is not used for model training and that its AI is hosted on European infrastructure; those are supplier statements that buyers should assess against their own requirements.
Our take: Permission design is the useful feature here: searching an inbox need not confer permission to send or delete messages. Start with a narrow read-only task, check the data-handling terms and expand access only when a workflow demonstrates its value. European hosting alone does not establish UK GDPR compliance.
Researchers find exposed cloud keys abused in under eight seconds
Researchers at Stony Brook University published 300 honeypot container images across eight public registries over four months, BleepingComputer reports. The fastest observed use of a planted AWS key was 7.86 seconds after publication on Docker Hub, and all 100 AWS keys were used within 15 minutes.
Exposed OpenAI and Anthropic keys were also used to exhaust test-account balances. The experiment found scanners could extract secrets directly from image layers without running the containers.
Our take: For businesses shipping AI applications, scanning after release is too late. Keep credentials out of build layers, use managed secrets at runtime and scan before publishing. If a key leaks, revoke it rather than merely deleting the visible file, and review usage and billing immediately.
