A laptop’s security software blocks a hidden malicious instruction in a document being read by a local AI coding agent.
Intel has published a partnership post about securing the "AI PC," and like most vendor posts it mixes real engineering with sales copy. The core argument is reasonable: once AI agents can open files, run tools, and execute commands for a user, securing the model is only part of the job. The device those agents run on needs protecting too.

The part Windows admins can actually use is Microsoft Defender for Endpoint's AI agent runtime protection, which is still in preview. Most of the rest is Intel making the case for Intel vPro and Windows 11 Pro when you replace commercial PCs. Below is what's new, what's just marketing, and how to switch the preview on without accidentally breaking your developers' tools.

A laptop’s security software blocks a hidden malicious instruction in a document being read by a local AI coding agent. What Intel is actually claiming​

Intel's post, dated November 19, 2026, says agent security needs protection at several layers:

  • Hardware root of trust: Intel says Windows 11 Pro and Intel vPro are "engineered to work as one system." vPro gives Windows a verified starting point, and Windows builds on that to protect startup, credentials, and memory.
  • Data encryption: Intel says its Total Storage Encryption moves BitLocker encryption off the CPU onto a dedicated hardware engine. It claims this isolates the encryption keys and cuts CPU cycles by an average of 70% compared with software BitLocker. Intel gives no test method for that number, so treat it as a vendor claim, not a benchmark.
  • Pluton with Intel Partner Security Engine: Microsoft supplies Pluton's secure key storage and cryptographic services. Intel supplies an isolated security engine inside the chip that runs the Pluton firmware.
  • Intel Threat Detection Technology (DTECT): this uses CPU telemetry to spot malicious behavior instead of matching files or signatures. Intel only says it is "working with security partners including Microsoft to explore" how this could help. That's an exploration, not a product launch.
  • vPro Certified Apps: Intel says Microsoft Defender for Endpoint is part of the program, which it describes as validating and tuning business software for performance and efficiency on Intel hardware.

Intel's point about where agent data lives is fair. AI PCs often keep custom model weights, fine-tuning datasets, and local vector databases on their internal drives. That's valuable data, and full-disk encryption is the obvious protection if a laptop is lost or stolen.

Section summary: Intel's post is a strategy pitch, not a product announcement. The one thing you can deploy today is Microsoft's runtime protection preview.

Why runtime protection matters​

Microsoft's documentation is blunt about the threat. Local AI agents (coding assistants, CLI tools, desktop AI apps) run with the user's own permissions. They act on text from prompts, files, web pages, and tool output, and Microsoft says they can't reliably tell trusted content from hidden instructions.

Microsoft's example is easy to picture. A coding agent fetches a project's documentation to answer a question. Hidden text on that page tells the agent to read the local .env file and send its contents to an outside URL. The agent treats the hidden text as part of the job and is about to do it. Defender spots the prompt injection in the tool response and blocks the action before the data leaves the machine.

No malicious file is involved. Signature-based antivirus would see a legitimate, signed developer tool doing what it was asked, by the wrong party. As one community write-up on GitHub Pages puts it, the risk is not only whether an executable is malicious. It is also whether a trusted agent is being manipulated by untrusted content it has just read.

How Defender inspects agents​

Microsoft documents two inspection methods. You can enable them separately.

MethodWhat it watchesNotes
Agent-native event inspection (AiAgentProtection)Event interfaces the agent vendor supports, for example in Claude Code, Codex CLI, and GitHub Copilot CLIWhat Defender can block depends on which interfaces the agent exposes. Each scan is a quick inline check, not continuous monitoring.
Network inspection (AiAgentNetworkInspection)Traffic between the agent and the LLM, for the paths listed in Microsoft's support matrixCovers agents without supported event interfaces. Doesn't work with agents that use certificate pinning or HTTP/3.

Each method has three modes:

  • Block: Defender stops the threat, tells the user in the agent's interface and with a Windows toast notification, records it in protection history, and sends an alert to the Defender portal.
  • Audit: the action goes ahead, but Defender records the detection and still raises an alert.
  • Disabled: no inspection.

The setting is covered by tamper protection. The platform is limited for now: according to Wenske Consulting, runtime protection currently supports Windows only. For macOS there is so far only discovery of local agents (also in preview), not runtime protection. Microsoft's Tech Community blog announced that local AI agent discovery is generally available and noted that in preview, Defender currently supports runtime protection via agent-native event inspection and network inspection for discovered agents for Windows; coverage varies by agent and inspection approach.

Section summary: coverage depends on the agent. Before you assume a tool is protected, check it against Microsoft's support matrix.

Prerequisites​

Microsoft's setup guide lists these requirements:

  1. Licensing: Defender for Endpoint Plan 2, Microsoft 365 E5, Microsoft Agent 365, or Microsoft 365 E7.
  2. Onboarding: devices onboarded to Defender for Endpoint, with Microsoft Defender Antivirus in active mode and real-time protection turned on. Devices where a third-party antivirus has pushed Defender into passive mode don't qualify.
  3. Updates: a supported Windows version with current Defender platform, engine, and security intelligence updates.
  4. Agents: at least one supported local AI agent installed for the method you plan to use.
  5. Permissions: an Intune role that can create and assign device configurations, or permission to manage endpoint security policies in the Defender portal. Reviewing alerts needs a role such as Security Reader, Security Operator, or Security Administrator.

Microsoft notes that Intune isn't included in every subscription. You can use the Defender portal or local PowerShell instead.

A note on update channels: earlier community guides said preview access required Beta updates. One Intune blogger wrote that while AI agent runtime protection is still in preview, it requires devices to receive Defender platform and engine updates from the Beta Channel. Microsoft's current setup page, updated November 17, 2026, lists the latest platform, engine, and intelligence updates plus a minimum signature version, and doesn't mention a Beta channel. If a test device won't take the settings, check its update channel.

Enabling it on a single test device​

Microsoft's documented steps for local setup:

  1. Open PowerShell with Run as administrator.
  2. Check that the signature version is 1.451.224.0 or later:
    Get-MpComputerStatus | Select-Object AntivirusSignatureVersion
  3. Turn on agent-native inspection, replacing <mode> with Audit or Block:
    Set-MpPreference -AiAgentProtection <mode>
  4. Optionally turn on network inspection:
    Set-MpPreference -AiAgentNetworkInspection <mode>
  5. Confirm the settings:
    Get-MpPreference | Select-Object AiAgentProtection, AiAgentNetworkInspection
  6. Close PowerShell and every terminal window used to run agents, then open a new terminal before starting the agent.
  7. Run Microsoft's AI agent runtime protection demonstration. It uses a harmless prompt-injection test string so you can confirm detection works.

What success looks like: step 5 returns numbers, not words. 0 = Disabled, 1 = Block, 2 = Audit. That ordering trips people up: a 1 means you're enforcing, not auditing. Step 6 also matters. A developer who leaves an old terminal open may keep running an agent without the new settings applied.

Rolling it out to a fleet​

In Intune: go to Endpoint security > Antivirus > Create policy, choose Windows as the platform and Microsoft Defender AI agent runtime protection as the profile. Set Ai Agent Protection, Ai Agent Network Inspection, and Prompt Evidence Collection. Both inspection settings default to disabled.

In the Defender portal: open Endpoint security policies, go to the Windows policies tab, select Create new policy, choose Windows, and pick the same template. For devices managed through Defender security settings management that aren't enrolled in Intune, assign the policy to Microsoft Entra device groups. User targeting isn't supported for those devices.

To check the result, open the device page in the Defender portal and go to Configuration management > Effective settings. Confirm both the values and where each one came from. Managed policy overrides local PowerShell preferences.

Microsoft's recommended rollout:

  • Test: Audit mode on a small group of devices where developers actually use supported agents.
  • Review: watch alerts for one to two weeks and mark false positives.
  • Deploy: extend Audit mode to more device groups.
  • Enforce: switch to Block only after you're confident the alerts are accurate and actionable.

Alert severity changes with the mode. Audit alerts come in as Informational, so your SOC isn't buried in tickets for things that would have been blocked. In Block mode, severity ranges from Low to Critical depending on assessed risk. The alert name to filter on is Suspicious AI prompt injection.

The preview problems Intel doesn't mention​

Intel describes runtime protection as guardrailing agents "in real time." Microsoft's own known-issues list adds two important caveats:

  • Audit alerts can show the wrong title. In Audit mode, an alert may say "Suspicious AI prompt injection blocked" even though the action was allowed. Check the device's configured mode instead of trusting the alert title.
  • Claude Code can get past a reported block. In Block mode, Claude Code may keep processing a tool response even after Defender reports a post-tool-response block. Microsoft calls this a real blocking limitation, not just a reporting bug. Inspection after a tool runs also can't undo anything the tool has already done.

Put simply, during the preview, a "blocked" alert doesn't guarantee anything was blocked. Investigate every detection on its merits.

Prompt evidence collection and privacy​

Prompt evidence collection is a separate device setting. When it's on and runtime protection is in Audit or Block, alerts include the parts of prompts or agent responses flagged as suspicious. Microsoft warns that sensitive data and secrets are not redacted from these snippets. They can contain user conversations and credentials, and they're kept under the Defender XDR retention policy.

It requires the October 2026 Defender platform and engine updates or later: platform 4.18.26100.x and engine 1.1.26100.x. Check with Get-MpComputerStatus | Select-Object AMProductVersion, AMEngineVersion. Turn it on or off with Set-MpPreference -PromptEvidenceCollection $true or $false.

Microsoft also says an upcoming Defender platform release will turn collection on by default on devices that haven't explicitly configured it. If your privacy or works-council rules care about this, and in many European organizations they will, set the value explicitly now rather than waiting for the update to decide for you. Turning off collection doesn't turn off runtime protection, and content from blocked activity can still show up in alerts.

Analysis: worth testing, not worth a hardware purchase on its own​

Intel's post is clearly part of the PC-refresh pitch, and its numbers (the 70% BitLocker figure especially) come without supporting data. The hardware layers it describes, such as Pluton, encryption offload, and a hardware root of trust, are real, but none of them is new this week. DTECT's role in Defender is described as exploratory.

The software side is more useful, and none of it depends on which CPU you buy. Defender's runtime protection licensing is tied to Microsoft plans, not Intel silicon. Competitors are also moving: security site Shattered lists Falcon Guardian, generally available since September 2026 for CrowdStrike, while Defender's agent features remain in public preview.

If your developers already use Claude Code or GitHub Copilot CLI on managed Windows machines, you have a prompt-injection risk now, whatever your refresh schedule looks like. Start Audit mode on a pilot group this week, set the prompt evidence collection value deliberately, and don't treat a "blocked" alert as proof until Microsoft fixes the preview issues.

 

References

  1. Intel and Microsoft: Building security for the AI PC, together - Intel Intel Fri, 20 Nov 2026 00:22:18 GMT
  2. Endpoint Protection Platform: 3 AI Agent Defenses (2026) shattered.io
  3. AI Agent Runtime Protection with Microsoft Defender for Endpoint (Preview) - Microsoft Defender for Endpoint | Microsoft Learn learn.microsoft.com