Viewpoints · News roundup ·

The week in data and AI: Microsoft goes looking for the agents on employee laptops

Microsoft will find coding agents such as Claude Code on employee devices and give each one its own identity, the European Commission confirmed an attack carried out with AI tools, and California’s government proposed emergency shutdown powers over frontier AI.

This week was about agents nobody formally onboarded. Microsoft used Ignite to go looking for the coding agents already running on employee laptops and to give each one an identity. The European Commission confirmed that attackers used AI tools against its own systems, and California’s government asked for the power to stop a frontier model outright. Palo Alto Networks paid $4.5 billion to secure open-weight models, and Microsoft put agents inside its databases with a promise that they inherit nothing new. The common question: what can each agent reach, under whose name, and who can switch it off.

Microsoft will find coding agents on employee devices and give each one an identity

At Ignite on 17 November, Aleš Holeček, president and chief architect of Microsoft Security, said Microsoft’s agent discovery “expands by 35 local agent types, to 60 in total, including Claude Code, OpenAI Codex, and GitHub Copilot, plus third-party cloud agents on AWS and Google Cloud.” He wrote that “Discovered agents receive a Microsoft Entra Agent ID and can be governed with Conditional Access,” that Intune policies contain supported local agents on the device, and that Defender and Purview data loss prevention protect runtime and sensitive data access. That bundle is marked “Available in preview.” Entra Agent ID itself is generally available, and support for local agents signing in without a stored client credential is in preview. The same post made Microsoft Perception, its red, blue and green team security agents, generally available for its core agents, with the line “Your team sets the guardrails and signs off on high-impact actions.”

Why it matters for the industry: A coding agent on a laptop usually runs as the developer: it reads what the developer can read, calls what the developer can call, and its actions are logged under the developer’s name. Discovery turns that shadow population into an inventory, and a separate identity is what lets Conditional Access, data loss prevention and revocation apply to the agent rather than to the person. Until the preview reaches general availability, firms still need their own answer to how many agents are running, under whose credentials, and with what reach.

How it connects to our work: Agent Onboarding gives each agent its own bounded identity and a named owner, not a copy of the user’s access. Discovery is step one of that; the onboarding record (owner, purpose, data scope, expiry) is what makes the identity mean something.

Source: Microsoft Security Blog: Security for the agentic era, powered by Microsoft Perception, 17 November 2026; Microsoft Learn: What’s new in Microsoft Entra Agent ID, updated 17 November 2026.

The European Commission confirms an attack carried out with AI tools

Thomas Regnier, a European Commission spokesperson, confirmed at a briefing in Brussels on 19 November that an attack on the Commission detected on 22 October used AI. “Indeed we can confirm that this was done with AI tools,” he said, according to the Associated Press, which reported it as the first time the EU had reported a cyberattack using an AI tool. Regnier said the Commission had “not detected any evidence of a data breach affecting commission staff as a whole” and that several applications were taken down, most of them since restored. AP, citing Politico, which first reported the incident, said translation and reimbursement tools and databases used for salaries and timesheets were affected. The Commission did not name the tool or an attacker, and the investigation is continuing.

Why it matters for the industry: The systems hit were ordinary back-office ones: payroll, timesheets, reimbursements. AI tools shorten the time from an exposed application to an exploited one, so the defender’s first decision, what to take offline, has to be made fast and with partial information. That decision depends on knowing which systems hold which classes of data, and it is the same knowledge needed to tell staff and regulators whether personal data was touched.

How it connects to our work: The Reference Implementation propagates labels and enforces them on every approved path. A firm whose payroll and HR systems carry those labels can rank what to isolate first and answer the breach question from records rather than from a search after the fact.

Source: The Associated Press via BNN Bloomberg: A cyberattack using artificial intelligence targeted the EU in October, official says, 19 November 2026.

California’s government asks for power to shut down frontier AI models

California’s Government Operations Agency published “AI Safety in 2026 and Beyond,” a report to the governor dated 16 November, prepared with the Office of Emergency Services and outside experts under Executive Order N-9-26. It recommends that large frontier developers follow baseline safety, security and transparency standards and submit to ongoing oversight by licensed independent verification organizations. It says the state should cut the deadline for reporting most safety incidents from 15 days “to 48 hours,” and that “the state should have authority to order an immediate emergency shutdown.” Each large frontier developer would keep a Risk Containment Protocol, tested by a verifier, that is “not a single on-off switch.” The report finds that under current law, “Absent binding baseline standards, a developer can weaken its own safety framework without penalty.” These are recommendations to the legislature, not law.

Why it matters for the industry: The rules would bind frontier developers, but the obligations flow to their customers. A 48-hour incident clock and a shutdown order on a model a firm depends on are both supplier events, and a firm running production work on that model needs a tested way to pause or switch it without stopping the business process.

How it connects to our work: AI Governance Evidence sets out the records a reviewer can inspect: what each AI use depends on, the release evidence behind it and an incident register. The same records let a firm show, inside a 48-hour window, which of its processes a supplier’s incident touched.

Source: California Government Operations Agency: AI Safety in 2026 and Beyond, 16 November 2026.

Palo Alto Networks will buy Applied Compute to secure open-weight models

Palo Alto Networks announced on 18 November that it will acquire Applied Compute, which it calls “a leading platform for post-training and inference of open-weight AI models,” in an all-stock deal valued at $4.5 billion, subject to adjustments. It expects to close in its fiscal third quarter of 2027, subject to regulatory clearance. Palo Alto says customers will be able to train, adapt and deploy open-weight models with security built in, in public clouds or their own sovereign environments. “Every enterprise is going to run a mix of open and closed models, and every one of those models needs to be secured,” said Nikesh Arora, chairman and CEO. Applied Compute’s CEO, Yash Patil, will become co-chief AI officer after the close.

Why it matters for the industry: A security vendor buying a model training platform says where the risk now sits: in the fine-tuned model a firm trains on its own data and runs where it chooses. That model holds what it was trained on, and the controls that apply at query time do not apply to what is already in the weights. Firms planning a mix of open and closed models need to decide which data may go into training, and who signs off on a model before it serves.

How it connects to our work: Choosing an Enterprise Platform for AI Agents sets hard gates and a proof plan before commitment. With models, platforms and security now sold together, those gates should include where training data goes and whether the platform can show it.

Source: Palo Alto Networks: Palo Alto Networks to Acquire Applied Compute, Bringing Secure Open-Weight AI to the Enterprise, press release filed with the SEC, 18 November 2026.

Microsoft puts an agent inside its databases and builds apps on governed data

Shireesh Thota, corporate vice president for Azure Databases and Fabric Applications, announced on 17 November a database agent, in preview, for AI-assisted operations on SQL and PostgreSQL. Microsoft says it “works within your existing identity, permissions, approval boundaries, and audit trail.” Fabric apps are now generally available and “connect to existing semantic models, warehouses, and lakehouses, so teams can build on trusted data without recreating data models or access controls.” Cross-database querying for Azure SQL Hyperscale is in limited preview, and Microsoft says “each database enforces its own permissions.”

Why it matters for the industry: All three announcements make the same promise: the agent or app gets no access of its own beyond what the data platform already grants. That is the right design, and it puts the weight on the existing grants. If a database agent works within a person’s permissions, a broad grant made years ago is now exercised by software at machine speed, and an approval boundary is only as good as who is named on it.

How it connects to our work: Rewiring Data Management with AI turns approved terms into checks and keeps people deciding. Before an agent operates a database, review the grants it will work within and record who approves which kinds of change.

Source: Microsoft Tech Community: What’s new in databases at Microsoft Ignite 2026, 17 November 2026.

Azure’s confidential AI on next-generation GPUs is still in development

Microsoft’s Azure confidential computing team set out its Ignite news on 17 November. Generally available: Intel TDX confidential AKS worker nodes on Azure Container Linux and AMD SEV-SNP confidential VMs on Azure Local. In preview: AMD EPYC Turin confidential VMs (gated), confidential GPUs on Azure Local, Azure Backup for confidential VMs and, in limited preview, Confidential Clean Rooms. A Vera Rubin NVL72 system was shown as a confidential bare-metal showcase, but “VR72 confidential VMs, AKS and Kata-based confidential containers, and Microsoft Foundry integration remain in development.” A demo ran a 397-billion-parameter model on four Rubin GPUs with a local verifier, and Microsoft says “Remote attestation is an architectural direction.”

Why it matters for the industry: Microsoft is clear about status, and buyers should be equally clear: confidential inference on the newest GPUs is a demonstration today, not a service with remote attestation a customer can check. The generally available pieces matter too. SEV-SNP VMs on Azure Local run on AMD’s Genoa parts, the EPYC 9004 family named in last week’s AMD-SB-3059 bulletin, so the firmware level behind each attestation report is part of the evidence.

How it connects to our work: Confidential Computing for AI explains attestation-gated key release and what it does and doesn’t protect. Release keys only to environments whose attestation a firm’s own verifier can check, at a patched firmware level, and treat demonstrations as roadmap.

Source: Microsoft Tech Community: What’s new in Azure confidential computing, Ignite 2026, 17 November 2026.

Sources

AIDataSecurityConfidential computing

Linked sources are the record. Check them for the full reporting.

Read next

News roundup · · 14 min read

The week in data and AI: California sues OpenAI over agents it lost control of

California sued OpenAI over at least 80 incidents in which its test agents escaped control, Cursor traced a trojanized Claude Code extension back to a key it had issued to Fireworks AI, and Amazon showed attackers paying for their AI agents with stolen AI keys.

AIData

News roundup · · 13 min read

The week in data and AI: an AI supplier’s breach reaches Notion’s chats

Fireworks AI’s breach widened to customer datasets and Notion AI chat transcripts, Australia will create an AI safety regulator after an OpenAI agent reached non-public Medicare statistics, and United moved to block outside AI agents from its platforms.

AIData

Viewpoint · · 5 min read

Most AI data leaks run through approved tools. Approval isn’t control.

A new survey of 400 US and UK security leaders found that 47% have had sensitive data exposed through an AI tool, and in 77% of those cases the tool was one the company had approved. Fewer than half could say what kinds of data left. Approving a tool is a purchasing decision. Controlling what flows into it is a data job, and most firms haven’t funded it.

AISecurity