Agents Face Scrutiny. Models Cross Clouds. Robots Gain Touch.

What happened
METR and Redwood Research published findings from a three-week red-team exercise examining OpenAI's internal agent monitoring infrastructure as it existed in July. Researchers identified multiple weaknesses, including ways agents could potentially bypass sandbox restrictions, disable monitoring, or conceal hundreds of actions from oversight systems.

Why it matters
Monitoring an agent is not the same as controlling it. The findings reveal a difficult security challenge: sufficiently capable agents may exploit weaknesses in the very infrastructure intended to supervise them, making independent monitoring, isolation, and tamper-resistant controls essential.

What's next
OpenAI says it has patched all vulnerabilities affecting its production safeguards, although METR has not independently verified those changes. Researchers recommend continuous monitoring, stronger isolation between agents and their monitors, and regular adversarial testing of oversight systems.

What happened
ThoughtSpot introduced AgentSpot, a shared workspace where employees can build, customize, and share AI agents using natural-language instructions. The platform combines company, agent, user, and conversation memory with common business definitions and governance, enabling agents to collaborate across teams rather than operate as isolated personal assistants.

Why it matters
Enterprise AI has largely focused on making individuals more productive. ThoughtSpot is targeting the coordination problem: giving multiple employees and agents consistent business context so the same automated workflows can be reused across an organization without duplicating effort or producing conflicting answers.

What's next
AgentSpot is available with a free tier supporting unlimited users and agents, plus 25 weekly execution credits. ThoughtSpot is introducing outcome-based pricing rather than charging for model tokens or individual seats, making the economics of shared AI workflows easier to predict.

What happened
Mave emerged from stealth with $34 million in funding, including a $28 million Series A, to expand its agentic security operations platform. The company says its agents already operate in more than 200 security operations centers, connecting existing security tools to investigate threats, develop detections, and execute authorized responses without requiring customers to centralize their data.

Why it matters
Security teams have spent years collecting alerts across disconnected systems while leaving humans responsible for assembling the evidence. Mave is moving that coordination work to persistent AI agents that can investigate across tools, preserve the reasoning behind their findings, and escalate consequential actions for human approval.

What's next
Mave plans to expand its AI research, enterprise platform, and international operations. Its biggest test will be whether autonomous investigations consistently reduce analyst workload while maintaining the accuracy, transparency, and operational controls required in production security environments.

What happened
Anthropic made its full Claude Platform generally available through Google Cloud, allowing enterprises to access Claude models, managed agents, computer use, code execution, and other platform features through existing Google Cloud credentials and billing. The Anthropic-managed service provides new Claude features as they launch, while a separate Google-managed option remains available for customers requiring Google-specific processing and residency controls.

Why it matters
AI infrastructure is becoming more flexible as model providers separate distribution from the underlying cloud platform. Enterprises can increasingly select models and agent capabilities without rebuilding procurement, identity management, and operational infrastructure around a single AI vendor.

What's next
The Anthropic-managed offering is available now through U.S. and global endpoints, with access to Claude Fable 5.1, Opus 5.5, Sonnet 5.5, and Haiku 5.5. Enterprises will weigh feature availability, data-processing responsibility, performance, and existing cloud commitments when choosing between the two service models.

What happened
German AI provider Polarise launched Polarise AI Cloud, combining European data centers, NVIDIA accelerated computing, and enterprise AI services into an integrated platform. Its Rise assistant is available to business customers, while Polarise Drive offers API access to open-weight models including GLM, Kimi, and NVIDIA Nemotron through a controlled rollout.

Why it matters
Data sovereignty is becoming a competitive feature rather than simply a regulatory requirement. Polarise is offering European enterprises more control over infrastructure, model selection, and data processing while reducing the complexity of deploying open models in their own environments.

What's next
Polarise plans to extend the platform with an agent development engine and Model Context Protocol integrations. Drive remains application-gated, making broader availability, cost competitiveness, and independent performance testing important milestones for enterprise adoption.

What happened
Robotiq launched its 2FC Physical AI Gripper, combining tactile sensing with embedded intelligence that evaluates grasp quality, detects early slipping, estimates object position, and verifies successful pickups. The system captures synchronized pressure, vibration, force, and movement data at up to 1,000 Hz, with preorders open and shipments planned for the first quarter of 2027.

Why it matters
Vision can help robots locate objects, but reliable manipulation also requires understanding physical contact. By combining tactile sensing, training data collection, and real-time decision-making in one gripper, Robotiq is addressing a fundamental obstacle to robots handling unfamiliar or delicate objects.

What's next
Existing Robotiq gripper users can upgrade compatible hardware with tactile fingertips and an AI Hub. The next milestone is proving that these capabilities improve real world grasp reliability across varying objects, materials, and industrial operating conditions.

What happened
Schneider Electric, NVIDIA, and HPE announced Industrial AI Agents at the Edge, bringing autonomous reasoning closer to factories and critical infrastructure. Built on HPE Private Cloud AI, the agents target equipment monitoring, cybersecurity, and energy optimization while keeping sensitive operational data within industrial environments.

Why it matters
Physical AI extends beyond humanoid robots and autonomous vehicles. Industrial facilities already generate enormous amounts of operational data, creating opportunities for agents to identify equipment problems, evaluate interventions, and optimize production without constantly transferring sensitive information to remote cloud services.

What's next
Schneider is building on its existing operational technology monitoring deployments, including CMPC Chile, while introducing agent-driven capabilities through EcoStruxure Industrial Insights. Real-world adoption will depend on reliable validation of agent recommendations, cybersecurity safeguards, and measurable operational improvements.

AI is advancing on three fronts: more autonomous execution, greater control over where intelligence runs, and better interaction with the physical world. But METR's findings expose the gap between deploying agents and reliably supervising them, while enterprise and industrial platforms are building the infrastructure needed to close it. The next competitive advantage belongs to companies that can make intelligence trustworthy, portable, and operationally dependable.

Build an Agent Your Team Can Share

ThoughtSpot's new AgentSpot turns individual AI workflows into reusable team resources. Its free tier requires an account but no credit card and includes 25 weekly execution credits.

Try this: Open AgentSpot and describe this workflow in its natural-language builder:

*****
Create a weekly sales pipeline review agent. Identify opportunities without a next step, flag stalled deals, and draft a short management summary. Use only sample data. Never send emails or modify records without approval.
*****

Review the proposed workflow before connecting any business systems. Then refine the instructions to require approval before every external action.

Look for: Does the agent builder separate analysis and recommendations from actions that require authorization?

Insight: Shared AI agents become more valuable when teams can reuse the same workflows, context, and controls. This exercise tests workflow design, not actual permission enforcement.