AI agents are reaching beyond the chat window, bringing new questions about who controls the systems they can access. Google Home’s agent integrations and Mandiant’s security architecture put those boundaries in focus, while Virginia’s data-centre measures raise questions about the physical footprint of AI services.

The edition also explores machine forecasting as a product category and the role of proprietary data in ad-tech consolidation.

In this edition — jump to a story

🤖 Google Home gives third-party AI agents a control plane

Google has released an early-access MCP server that lets compatible assistants inspect Google Home structures, read current and historical device state, and execute supported actions. Access requires a Google Cloud project, OAuth credentials, and a Google Home Premium Advanced subscription; Google applies rate limits and blocks sensitive actions such as unlocking doors.

Why it matters: MCP is moving beyond enterprise software into the physical environment. Standardised tool access makes agents more useful and portable, but the consequences of a bad tool call now extend to cameras, appliances, and household telemetry. Consent, least privilege, revocation, and action-level policy become part of the product rather than deployment paperwork.

The Bigger Picture

The agent supply-chain risks explored in the previous briefing now extend beyond software and data. With Google Home, an agent’s external tools can also affect the physical environment. The same questions of provenance, permissions and oversight take on a new dimension when a tool call can change what happens inside a home.

☁️ Virginia makes data-centre growth a transparency and energy-policy issue

Virginia, home to the world’s largest concentration of data centres, announced a Data Center Accountability Framework. The plan includes a ban on non-disclosure agreements for projects of 25 MW or more and incentives for renewable backup power instead of natural gas or diesel; some measures still require legislative approval.

Why it matters: Cloud capacity is no longer negotiated only between providers, utilities, and large customers. Permitting, power costs, emissions, and disclosure are becoming constraints on where infrastructure can be built and how quickly it can come online. Capacity planning increasingly includes policy and community acceptance alongside land, chips, and electricity.

The Bigger Picture

The pressure on electricity networks and efforts to make data-centre demand more flexible are increasingly shaping public policy. Virginia’s framework illustrates a further shift: the AI infrastructure build-out is becoming a question of who bears its costs, who gets a say, and what operators must disclose before construction begins.

🛡️ Mandiant turns agent security into an architecture problem

Mandiant’s AI Risk and Resilience 2026 report argues that autonomous systems need dynamic, cryptographically bound identities instead of long-lived service accounts. It recommends separating compute identity from data access, requiring renewed human authorisation for high-risk actions, governing tools and RAG boundaries, treating coding assistants and MCP servers as privileged sessions, and feeding model, API, harness, and endpoint telemetry into real-time detection.

Why it matters: Agent security cannot stop at prompt filters. An agent combines identity, delegated authority, data, tools, models, APIs, and instructions that may change at runtime. Mandiant’s useful contribution is to connect familiar controls — workload identity, just-in-time secrets, segmentation, allowlists, SIEM/EDR telemetry, rate limits, and circuit breakers — into a defence-in-depth model for non-deterministic systems.

The Bigger Picture

Anthropic’s findings on autonomous attacks and AIR’s focus on agent dependencies expose different parts of the same security challenge. Mandiant brings them together from the defender’s side: as agents gain tools and authority, identity, authorisation and continuous monitoring must govern not only what an agent can access, but what it can do over time.

🚀 Mantic raises $25 million after AI forecasting beats human competitors

London-based Mantic raised a $25 million seed round led by Radical Ventures after its system outperformed human forecasters in the summer 2026 Metaculus Cup. The company specialises frontier models for probabilistic forecasting, tests them against historical outcomes, grades their performance, and iterates; Reuters reports early interest from companies, government agencies, hedge funds, and trading firms.

Why it matters: The important startup signal is not another general-purpose chatbot, but a narrow system evaluated against outcomes. Forecasting creates a measurable feedback loop: predictions carry probabilities and can be scored when events resolve. Commercial credibility will depend on sustained calibration across changing conditions, not a single tournament result.

📣 Infillion buys Foursquare to own more of the ad-tech data layer

Infillion is acquiring Foursquare, which will continue operating as an independent brand and selling data to customers including rival ad platforms. Axios reports that the deal adds a location database spanning more than 100 million points of interest and complements Infillion’s earlier purchase of Catalina, whose data connects advertising with consumer purchases.

Why it matters: Ad tech is consolidating around data that can link exposure to offline behaviour. Combining media buying, location signals, and purchase data may improve targeting and measurement, but it also concentrates control over attribution inside fewer platforms. Advertisers will need clarity on provenance, consent, interoperability, and whether reported outcomes can be independently verified.

Sources