Microsoft · Project Perception · Agent security
Microsoft decided the AI needs a supervisor. The supervisor is also an AI.
Microsoft’s October 19 expansion of Project Perception treats agents as potential insider risks: useful systems that may hold legitimate access while making mistakes, encountering manipulation or drifting beyond their assignments. The company describes task-scoped Dynamic Permissions, continuously updated Dynamic Agent Risk and Dynamic Conditional Access that can narrow authority as risk changes. Guardian Agents and playbooks would help investigate and coordinate responses. The architectural principle is external enforcement: the model should not control the boundary that determines what it may do.
An illustrative records assistant might need permission to read a folder without needing authority to rewrite it or send its contents elsewhere. If suspicious behavior emerges, a surrounding system could preserve safe reads while blocking writes, require human approval or stop execution. Our view is that a useful security response should be proportionate to the action and evidence, rather than treating every warning as a choice between unrestricted access and shutting down everything. The controls also need to remain effective if either the model or its harness behaves unexpectedly.
Microsoft says its internal agent deployments inform the design. That experience is relevant, but the announcement does not independently establish protection across every customer environment or representative failure case. A supervising agent can also misread evidence or receive excessive privileges. Its security assignment is a reason to constrain it carefully, rather than assume it is trustworthy. This is a direction worth evaluating through observed outcomes and independent testing. The agent has acquired a supervisor. Human Resources would still like to know who is responsible for both performance reviews.
Source: The expanded controls, Guardian Agents and external enforcement ↗Source: Microsoft’s broader argument for continuous agent supervision ↗NVIDIA · Modal · CPU infrastructure
The AI has a powerful model. Its ordinary computing chores still need a processor.
Modal’s October 19 report describes early tests of NVIDIA’s Arm-compatible Vera CPU for its Sandboxes service. Its Agentless Agent Bench uses fifty Terminal-Bench solutions and verifiers in prewarmed containers, removing network calls to isolate CPU work. Modal reports up to twice the task throughput of the tested Intel x86 server under heavy concurrent load. This is a workload-specific, company-run benchmark—not a demonstration that complete agents reason or finish every assignment twice as quickly.
Throughput counts completed tasks over time; latency measures how long an individual task takes. Modal reports that serial tests on a quiet machine showed less dramatic advantages, with Vera not winning every task. The stronger result appeared when many heterogeneous workloads competed for resources. That distinction explains why a processor can be interesting for a shared cloud service even if an isolated task does not suddenly run twice as fast. Our practical question is how well a host continues serving different users when their workloads become busy at the same time.
Modal says supporting Arm requires changes across its infrastructure and image-building system, and broader rollout remains ahead. Its benchmark is useful evidence to examine, while fleet reliability, software compatibility and customer economics still need evaluation. The connection to the security story is that an agent is an operating system of parts: model calls, tools, execution environments and the machines underneath them. Improving one part can help without proving an equivalent gain across the whole. The model may have a plan. Sometimes the plan is waiting for a CPU to finish its chores.
Source: Modal’s benchmark design, named comparison and rollout work ↗OVHcloud · Connected capacity · European AI
The industry is hunting for more computing space. OVHcloud says it has a vacancy.
OVHcloud chief executive Octave Klaba told Reuters on October 20 that the company has 250 megawatts of connected capacity, with 140 MW operating and 110 MW available for new business. He argues that this lets OVHcloud add customers without first constructing more facilities. That is available connected capacity, not proof that every remaining megawatt already has all the accelerators, software and customers needed for an AI workload. The disclosed numbers also make “about half” an approximation, rather than exactly half.
The contrast with Sunday’s reported British power delay is useful. Access to connected infrastructure can change a provider’s starting position, but a viable service still needs the right equipment and demand. Our view is that building capacity and making productive use of it are distinct achievements. A customer needs a suitable service on a dependable timetable; unused capacity is an opportunity to deliver one, rather than delivery already accomplished. The same distinction applies when evaluating ambitious plans for additional public or private computing facilities.
Klaba says an OVHai Workspace pilot is due shortly and OVHcloud’s first AI model is planned for early 2027, using a mixture-of-experts design. Those are prospective milestones with performance still to demonstrate. He also criticizes the customer demand underpinning Europe’s state-backed computing projects; that is the perspective of a commercial participant, rather than a settled verdict on every public objective. The company must still turn its infrastructure position into competitive services. The future needs electricity, software and customers. It also appears to need a landlord with something already connected.
Source: Reuters on connected capacity, the proposed model and Klaba’s comments ↗