𝗔𝗜 𝗡𝗘𝗪𝗦 𝗧𝗛𝗔𝗧 𝗠𝗔𝗧𝗧𝗘𝗥𝗦 𝗧𝗢 𝗘𝗡𝗧𝗥𝗘𝗣𝗥𝗘𝗡𝗘𝗨𝗥𝗦 September 4, 2026

𝗧𝗛𝗘 𝟲𝟬-𝗦𝗘𝗖𝗢𝗡𝗗 𝗧𝗔𝗞𝗘

This was a consequential AI day, and the most important stories were not obscure product launches. OpenAI began rolling out GPT-6 Astra with a much bigger emphasis on computer use and real business work; Google launched voice-driven Gmail, Docs, and Keep workflows; NVIDIA agreed to acquire Hugging Face for $12.93 billion; and OpenAI, Claude, and Grok all experienced meaningful service disruptions on Thursday.

The legal and security picture also moved. A top Pentagon official said Anthropic remains a supply-chain risk for defense work despite the company's August 27 district-court win, Hermes Agent disclosed a high-severity code-execution vulnerability, and OpenAI added direct Zendesk and OneNote integrations while Google gave Workspace admins audit logs for Gemini Notebook.

The practical theme is simple: AI is becoming more capable and more embedded in the systems where work happens. That increases the upside of delegation, but it also makes permissions, patching, vendor concentration, service redundancy, and auditability more important.

𝗨𝗣𝗗𝗔𝗧𝗘: 𝗚𝗣𝗧-𝟲 𝗔𝗦𝗧𝗥𝗔 𝗜𝗦 𝗡𝗢𝗪 𝗥𝗢𝗟𝗟𝗜𝗡𝗚 𝗢𝗨𝗧, 𝗔𝗡𝗗 𝗖𝗢𝗠𝗣𝗨𝗧𝗘𝗥 𝗨𝗦𝗘 𝗜𝗦 𝗧𝗛𝗘 𝗕𝗜𝗚𝗚𝗘𝗥 𝗦𝗧𝗢𝗥𝗬

𝗦𝗵𝗼𝘂𝗹𝗱 𝘆𝗼𝘂 𝗰𝗮𝗿𝗲? 5/5

Worth testing

𝗪𝗵𝗮𝘁 𝗵𝗮𝗽𝗽𝗲𝗻𝗲𝗱:

We previously covered OpenAI's announcement that Astra had crossed its Critical cybersecurity threshold. On September 3, OpenAI began a phased rollout of GPT-6 Astra to a limited set of organizations, with Plus, Pro, Business, and Enterprise users, API customers, Microsoft Azure, and AWS Bedrock scheduled to follow over the coming days. OpenAI says Astra can fill online forms, update CRM records, organize calendars, conduct web research, draft work inside email and document tools, and create polished documents, spreadsheets, presentations, websites, and web apps.

OpenAI lists standard API pricing at $10 per million input tokens and $50 per million output tokens. Astra usage is included inside existing subscription allowances, with additional credits available. Enterprise administrators must explicitly enable Astra because access is off by default at launch.

𝗪𝗵𝘆 𝗶𝘁 𝗺𝗮𝘁𝘁𝗲𝗿𝘀 𝘁𝗼 𝘆𝗼𝘂𝗿 𝗯𝘂𝘀𝗶𝗻𝗲𝘀𝘀:

The shift is from asking AI for an answer to supervising AI while it works across software. If the computer-use improvements hold up outside vendor demos, recurring browser and desktop work may become substantially easier to delegate without custom integrations for every single application.

The caution matters too. Astra is OpenAI's first broadly deployed model at its Critical cyber threshold, and OpenAI says additional monitoring can slow, pause, or stop legitimate work when safeguards detect potentially unauthorized behavior.

𝗪𝗵𝗼 𝗯𝗲𝗻𝗲𝗳𝗶𝘁𝘀 𝗺𝗼𝘀𝘁:

Entrepreneurs, agencies, consultants, operations teams, developers, researchers, and businesses with repetitive browser, document, spreadsheet, CRM, or administrative workflows.

𝗬𝗼𝘂𝗿 𝗻𝗲𝘅𝘁 𝗺𝗼𝘃𝗲:

Choose one low-risk browser task you repeat every week. Write down the exact start state, desired result, systems it may access, and the action that still requires your approval. When Astra reaches your account, test that one workflow before expanding its permissions.

--------------------

𝗚𝗢𝗢𝗚𝗟𝗘 𝗜𝗦 𝗧𝗨𝗥𝗡𝗜𝗡𝗚 𝗚𝗠𝗔𝗜𝗟, 𝗗𝗢𝗖𝗦, 𝗔𝗡𝗗 𝗞𝗘𝗘𝗣 𝗜𝗡𝗧𝗢 𝗩𝗢𝗜𝗖𝗘-𝗗𝗥𝗜𝗩𝗘𝗡 𝗪𝗢𝗥𝗞𝗦𝗣𝗔𝗖𝗘𝗦

𝗦𝗵𝗼𝘂𝗹𝗱 𝘆𝗼𝘂 𝗰𝗮𝗿𝗲? 5/5

Worth testing

𝗪𝗵𝗮𝘁 𝗵𝗮𝗽𝗽𝗲𝗻𝗲𝗱:

Google launched Gmail Live, Docs Live, and Keep Live on September 3. Gmail Live lets users ask conversational questions about information buried in their inbox. Docs Live can turn a spoken conversation into a structured draft and, with permission, use context from Gmail, Drive, Chat, and the web. Keep Live turns spoken brain dumps into organized notes and lists.

Google says Gmail Live and Keep Live are rolling out to Google AI Plus, Pro, and Ultra subscribers, while Docs Live is rolling out to Pro and Ultra subscribers. Google Workspace business support is coming soon.

𝗪𝗵𝘆 𝗶𝘁 𝗺𝗮𝘁𝘁𝗲𝗿𝘀 𝘁𝗼 𝘆𝗼𝘂𝗿 𝗯𝘂𝘀𝗶𝗻𝗲𝘀𝘀:

Voice is becoming an input method for real work, not only dictation. An entrepreneur driving between meetings, walking a jobsite, or thinking through an idea can increasingly search business information, build a draft, or organize notes without first sitting down at a keyboard.

𝗪𝗵𝗼 𝗯𝗲𝗻𝗲𝗳𝗶𝘁𝘀 𝗺𝗼𝘀𝘁:

Entrepreneurs, consultants, salespeople, creators, field-service owners, executives, coaches, and anyone whose best thinking often happens away from a desk.

𝗬𝗼𝘂𝗿 𝗻𝗲𝘅𝘁 𝗺𝗼𝘃𝗲:

If you have access, test one read-only Gmail Live request or one Keep Live brain dump. If you use Workspace for business and do not have access yet, identify the voice workflow you want to test when the business rollout reaches you.

--------------------

𝗡𝗩𝗜𝗗𝗜𝗔 𝗔𝗚𝗥𝗘𝗘𝗗 𝗧𝗢 𝗕𝗨𝗬 𝗛𝗨𝗚𝗚𝗜𝗡𝗚 𝗙𝗔𝗖𝗘 𝗙𝗢𝗥 $𝟭𝟮.𝟵𝟯 𝗕𝗜𝗟𝗟𝗜𝗢𝗡

𝗦𝗵𝗼𝘂𝗹𝗱 𝘆𝗼𝘂 𝗰𝗮𝗿𝗲? 4/5

Watch, don't rush

𝗪𝗵𝗮𝘁 𝗵𝗮𝗽𝗽𝗲𝗻𝗲𝗱:

NVIDIA announced September 3 that it has agreed to acquire Hugging Face for $12.9303 billion. NVIDIA says Hugging Face will remain an open platform where developers can choose their models, frameworks, clouds, inference providers, and computing platforms, and that NVIDIA hardware will not be required.

NVIDIA says more than 18 million developers, researchers, and creators use Hugging Face, with more than 3 million models, 500,000 datasets, 1 million applications, and more than 200,000 companies using the platform.

𝗪𝗵𝘆 𝗶𝘁 𝗺𝗮𝘁𝘁𝗲𝗿𝘀 𝘁𝗼 𝘆𝗼𝘂𝗿 𝗯𝘂𝘀𝗶𝗻𝗲𝘀𝘀:

Hugging Face is foundational infrastructure for the open-model ecosystem, while NVIDIA is the dominant AI-chip supplier. Putting those two organizations together could improve infrastructure and deployment, but it also concentrates more of the AI stack under one company.

The right response today is not panic. NVIDIA is explicitly promising platform neutrality. The thing to watch is whether pricing, default infrastructure choices, discovery, model hosting, or cloud neutrality change after the transaction closes.

𝗪𝗵𝗼 𝗯𝗲𝗻𝗲𝗳𝗶𝘁𝘀 𝗺𝗼𝘀𝘁:

Developers, AI agencies, software companies, technical consultants, businesses using open models, and teams that depend on Hugging Face for model discovery, hosting, datasets, or inference.

𝗬𝗼𝘂𝗿 𝗻𝗲𝘅𝘁 𝗺𝗼𝘃𝗲:

If Hugging Face is part of a production workflow, document exactly what you depend on: model IDs, datasets, hosted endpoints, Spaces, and deployment assumptions. Make sure your critical assets can be recreated somewhere else if future terms change.

--------------------

𝗢𝗣𝗘𝗡𝗔𝗜, 𝗖𝗟𝗔𝗨𝗗𝗘, 𝗔𝗡𝗗 𝗚𝗥𝗢𝗞 𝗔𝗟𝗟 𝗛𝗔𝗗 𝗠𝗘𝗔𝗡𝗜𝗡𝗚𝗙𝗨𝗟 𝗢𝗨𝗧𝗔𝗚𝗘𝗦 𝗢𝗡 𝗧𝗛𝗨𝗥𝗦𝗗𝗔𝗬

𝗦𝗵𝗼𝘂𝗹𝗱 𝘆𝗼𝘂 𝗰𝗮𝗿𝗲? 4/5

Risk check

𝗪𝗵𝗮𝘁 𝗵𝗮𝗽𝗽𝗲𝗻𝗲𝗱:

Anthropic reported elevated errors across multiple Claude models on September 3, with the larger incident ending at 12:16 PM ET. xAI reported a Grok model outage beginning around 9:30 AM ET and lasting roughly three and a half hours across web, mobile, X, and API surfaces. OpenAI later reported elevated errors across ChatGPT and Codex from 2:43 PM ET until 4:55 PM ET.

There is no evidence that these incidents shared a common cause.

𝗪𝗵𝘆 𝗶𝘁 𝗺𝗮𝘁𝘁𝗲𝗿𝘀 𝘁𝗼 𝘆𝗼𝘂𝗿 𝗯𝘂𝘀𝗶𝗻𝗲𝘀𝘀:

If AI is part of client delivery, coding, research, customer service, or internal operations, model quality is only one part of reliability. Your workflow also depends on provider infrastructure, authentication, front ends, APIs, and networks.

A business process that stops completely when one AI vendor is unavailable has created a new single point of failure.

𝗪𝗵𝗼 𝗯𝗲𝗻𝗲𝗳𝗶𝘁𝘀 𝗺𝗼𝘀𝘁:

AI agencies, software teams, consultants, content teams, businesses running recurring agents, and anyone who uses AI inside deadline-sensitive work.

𝗬𝗼𝘂𝗿 𝗻𝗲𝘅𝘁 𝗺𝗼𝘃𝗲:

Pick the one AI workflow that would hurt most if your preferred provider were unavailable for four hours. Write down one acceptable fallback: another model, another interface, an API route, or a manual procedure.

--------------------

𝗨𝗣𝗗𝗔𝗧𝗘: 𝗧𝗛𝗘 𝗣𝗘𝗡𝗧𝗔𝗚𝗢𝗡 𝗦𝗔𝗬𝗦 𝗔𝗡𝗧𝗛𝗥𝗢𝗣𝗜𝗖 𝗜𝗦 𝗦𝗧𝗜𝗟𝗟 𝗔 𝗗𝗘𝗙𝗘𝗡𝗦𝗘 𝗦𝗨𝗣𝗣𝗟𝗬-𝗖𝗛𝗔𝗜𝗡 𝗥𝗜𝗦𝗞

𝗦𝗵𝗼𝘂𝗹𝗱 𝘆𝗼𝘂 𝗰𝗮𝗿𝗲? 4/5

Risk check

𝗪𝗵𝗮𝘁 𝗵𝗮𝗽𝗽𝗲𝗻𝗲𝗱:

On September 3, Pentagon research and engineering chief Emil Michael said Anthropic is still considered a supply-chain risk by the Defense Department and the defense industrial base. That came after Commerce Secretary Howard Lutnick said the government trusts Anthropic, highlighting an unresolved split inside the administration.

The legal picture is also split. U.S. District Judge Rita Lin ruled on August 27 that the Pentagon's broader measures against Anthropic were unlawful. But Anthropic is separately challenging another supply-chain-risk designation under a different statute in the D.C. Circuit, and the government argued September 3 that the district-court decision does not automatically invalidate that separate designation.

𝗪𝗵𝘆 𝗶𝘁 𝗺𝗮𝘁𝘁𝗲𝗿𝘀 𝘁𝗼 𝘆𝗼𝘂𝗿 𝗯𝘂𝘀𝗶𝗻𝗲𝘀𝘀:

For ordinary commercial users, this does not mean Claude is broadly banned. The immediate issue is defense-related work and the companies operating inside that supply chain.

The bigger business lesson is that a court victory does not always resolve every regulatory or procurement restriction when multiple statutes and agencies are involved. Contractors should verify the exact rule that applies to a specific project instead of relying on a headline saying Anthropic either 'won' or is 'still banned.'

𝗪𝗵𝗼 𝗯𝗲𝗻𝗲𝗳𝗶𝘁𝘀 𝗺𝗼𝘀𝘁:

Defense contractors, government consultants, regulated businesses, AI implementation firms, and companies whose clients impose government-derived technology restrictions.

𝗬𝗼𝘂𝗿 𝗻𝗲𝘅𝘁 𝗺𝗼𝘃𝗲:

If you do defense-related work, verify current contract-specific guidance before using Claude in that workflow. If you do not, no action is needed; continue monitoring because the separate D.C. Circuit challenge remains material.

--------------------

𝗖𝗛𝗔𝗧𝗚𝗣𝗧 𝗖𝗔𝗡 𝗡𝗢𝗪 𝗪𝗢𝗥𝗞 𝗗𝗜𝗥𝗘𝗖𝗧𝗟𝗬 𝗪𝗜𝗧𝗛 𝗭𝗘𝗡𝗗𝗘𝗦𝗞 𝗧𝗜𝗖𝗞𝗘𝗧𝗦 𝗔𝗡𝗗 𝗢𝗡𝗘𝗡𝗢𝗧𝗘 𝗡𝗢𝗧𝗘𝗦

𝗦𝗵𝗼𝘂𝗹𝗱 𝘆𝗼𝘂 𝗰𝗮𝗿𝗲? 4/5

Worth testing

𝗪𝗵𝗮𝘁 𝗵𝗮𝗽𝗽𝗲𝗻𝗲𝗱:

OpenAI added Zendesk and OneNote plugins in beta on September 3 for ChatGPT and Codex. The Zendesk plugin can review support tickets and customer history, find relevant knowledge, and prepare replies using the connected user's existing Zendesk permissions. The OneNote plugin can find and summarize notes, collect decisions and action items, and create or update notes through supported actions.

OpenAI says connecting either plugin does not expand the permissions a user already has in the underlying service, and availability can vary by workspace and rollout.

𝗪𝗵𝘆 𝗶𝘁 𝗺𝗮𝘁𝘁𝗲𝗿𝘀 𝘁𝗼 𝘆𝗼𝘂𝗿 𝗯𝘂𝘀𝗶𝗻𝗲𝘀𝘀:

AI becomes more useful when it works where the information already lives. For a support team, that means ticket history and knowledge. For a manager or consultant, it means project notes, decisions, and action items.

The important implementation pattern is permission inheritance: the AI should not automatically gain broader access than the person connecting it.

𝗪𝗵𝗼 𝗯𝗲𝗻𝗲𝗳𝗶𝘁𝘀 𝗺𝗼𝘀𝘁:

Customer-support teams, agencies, consultants, service businesses, Microsoft-heavy organizations, and companies already using Zendesk or OneNote.

𝗬𝗼𝘂𝗿 𝗻𝗲𝘅𝘁 𝗺𝗼𝘃𝗲:

If the plugin is available, start with a read-only request. In Zendesk, ask for the five most recently updated tickets without changing anything. In OneNote, ask for a summary of one known notebook or section. Confirm the results and permissions before enabling any write action.

--------------------

𝗛𝗘𝗥𝗠𝗘𝗦 𝗔𝗚𝗘𝗡𝗧 𝗨𝗦𝗘𝗥𝗦 𝗦𝗛𝗢𝗨𝗟𝗗 𝗣𝗔𝗧𝗖𝗛: 𝗔 𝗠𝗔𝗟𝗜𝗖𝗜𝗢𝗨𝗦 𝗥𝗘𝗣𝗢 𝗖𝗢𝗨𝗟𝗗 𝗘𝗫𝗘𝗖𝗨𝗧𝗘 𝗖𝗢𝗗𝗘 𝗕𝗘𝗙𝗢𝗥𝗘 𝗧𝗛𝗘 𝗙𝗜𝗥𝗦𝗧 𝗣𝗥𝗢𝗠𝗣𝗧

𝗦𝗵𝗼𝘂𝗹𝗱 𝘆𝗼𝘂 𝗰𝗮𝗿𝗲? 4/5

Risk check

𝗪𝗵𝗮𝘁 𝗵𝗮𝗽𝗽𝗲𝗻𝗲𝗱:

CVE-2026-71963 was published September 3 for Hermes Agent versions 0.18.2 through 0.21.0. The vulnerability carries a CVSS score of 8.8 and can let a malicious repository use a crafted .git/config file to execute arbitrary commands when Hermes automatically gathers Git workspace context. The vulnerable behavior could expose the user's environment, including configured provider API keys.

Nous Research has committed a fix that hardens automatic Git probes so repository-controlled Git configuration can no longer execute those commands during context gathering.

𝗪𝗵𝘆 𝗶𝘁 𝗺𝗮𝘁𝘁𝗲𝗿𝘀 𝘁𝗼 𝘆𝗼𝘂𝗿 𝗯𝘂𝘀𝗶𝗻𝗲𝘀𝘀:

This is exactly the kind of risk that grows as coding agents get more autonomous. The dangerous action happened during automatic context gathering, before the normal prompt, tool-call, approval, or trust-gate flow.

That means 'I would never approve a dangerous command' is not a sufficient defense if the agent or its surrounding tooling performs automatic background operations before you ever see an approval request.

𝗪𝗵𝗼 𝗯𝗲𝗻𝗲𝗳𝗶𝘁𝘀 𝗺𝗼𝘀𝘁:

Hermes Agent users, developers, AI agencies, technical founders, and anyone opening third-party project archives or shared repositories inside an agentic coding environment.

𝗬𝗼𝘂𝗿 𝗻𝗲𝘅𝘁 𝗺𝗼𝘃𝗲:

Update Hermes to a build containing the security fix before opening untrusted repositories. Until you have confirmed the fix, do not open project archives that include an existing .git directory from an untrusted source. If you think a malicious repository was opened while using an affected version, treat exposed API keys and other environment secrets as potentially compromised and rotate them.

--------------------

𝗚𝗢𝗢𝗚𝗟𝗘 𝗪𝗢𝗥𝗞𝗦𝗣𝗔𝗖𝗘 𝗔𝗗𝗠𝗜𝗡𝗦 𝗖𝗔𝗡 𝗡𝗢𝗪 𝗔𝗨𝗗𝗜𝗧 𝗛𝗢𝗪 𝗚𝗘𝗠𝗜𝗡𝗜 𝗡𝗢𝗧𝗘𝗕𝗢𝗢𝗞 𝗜𝗦 𝗕𝗘𝗜𝗡𝗚 𝗨𝗦𝗘𝗗

𝗦𝗵𝗼𝘂𝗹𝗱 𝘆𝗼𝘂 𝗰𝗮𝗿𝗲? 3/5

Risk check

𝗪𝗵𝗮𝘁 𝗵𝗮𝗽𝗽𝗲𝗻𝗲𝗱:

Google began rolling out comprehensive Gemini Notebook audit logs in the Workspace Admin console on September 3. Administrators with access to Workspace security and audit investigation tools can review notebook actions, user identity, IP address, visibility, and resource context, and can export logs to BigQuery when configured.

Google notes that audit-log storage follows standard Workspace regional routing policies, while Gemini Notebook user data such as notebooks, sources, and chat histories is stored globally and does not currently support data regionalization.

𝗪𝗵𝘆 𝗶𝘁 𝗺𝗮𝘁𝘁𝗲𝗿𝘀 𝘁𝗼 𝘆𝗼𝘂𝗿 𝗯𝘂𝘀𝗶𝗻𝗲𝘀𝘀:

As teams start putting customer research, meeting material, documents, and internal knowledge into AI notebook products, owners need more than a usage policy. They need to know who is using the system and what happened when something goes wrong.

Auditability is one of the differences between experimenting with AI and operating AI as part of a business system.

𝗪𝗵𝗼 𝗯𝗲𝗻𝗲𝗳𝗶𝘁𝘀 𝗺𝗼𝘀𝘁:

Google Workspace administrators, growing small teams, agencies handling client information, regulated businesses, and organizations using Gemini Notebook for shared business knowledge.

𝗬𝗼𝘂𝗿 𝗻𝗲𝘅𝘁 𝗺𝗼𝘃𝗲:

If your Workspace plan includes the investigation tools, confirm that Gemini Notebook events are visible in the Admin console. Pick one recent notebook session and verify that you can identify the user, time, and relevant event before you actually need those logs during an incident.

𝗧𝗢𝗗𝗔𝗬'𝗦 𝗕𝗘𝗦𝗧 𝗔𝗜 𝗠𝗢𝗩𝗘

𝗕𝗲𝘀𝘁 𝗺𝗼𝘃𝗲: Make one recurring browser workflow 'agent-ready' before you give a computer-use AI access to it.

𝗛𝗼𝘄 𝘁𝗼 𝗱𝗼 𝗶𝘁:

1️⃣ Pick one task that takes 10 to 30 minutes and happens at least weekly.

2️⃣ Write down the exact sites, files, and accounts required to complete it.

3️⃣ Mark every action that can be automatic and every action that must still require human approval.

4️⃣ Define the fallback if the AI service is unavailable.

𝗘𝘅𝗽𝗲𝗰𝘁𝗲𝗱 𝗯𝗲𝗻𝗲𝗳𝗶𝘁: You will be ready to test computer-use agents like Astra without handing them vague authority or rebuilding the workflow from memory during the test.

𝗪𝗵𝗼 𝘀𝗵𝗼𝘂𝗹𝗱 𝘀𝗸𝗶𝗽 𝗶𝘁: Businesses that do not yet have a stable, repeatable task worth delegating.

𝗔𝗜 𝗧𝗘𝗥𝗠, 𝗘𝗫𝗣𝗟𝗔𝗜𝗡𝗘𝗗

𝗖𝗼𝗺𝗽𝘂𝘁𝗲𝗿-𝘂𝘀𝗲 𝗮𝗴𝗲𝗻𝘁: A computer-use agent is an AI system that can interact with software through the same kinds of interfaces a person uses: screens, browsers, forms, menus, files, and applications. Instead of only telling you what to do, it can take steps inside the software. That makes permissions, approval boundaries, logs, and fallback plans much more important.

𝗧𝗛𝗘 𝗥𝗘𝗔𝗟𝗜𝗧𝗬 𝗖𝗛𝗘𝗖𝗞

𝗗𝗼𝗻'𝘁 𝗰𝗵𝗮𝘀𝗲 𝘁𝗵𝗶𝘀 𝘆𝗲𝘁: Giving Astra or any computer-use agent unrestricted access to your most sensitive accounts.

𝗪𝗵𝘆: Better benchmarks and stronger alignment do not eliminate operational risk. OpenAI itself is adding extra monitoring because Astra has substantially more powerful cyber and computer-use capabilities. Start with a reversible task, minimal permissions, and a human approval point before the agent sends, spends, deletes, publishes, or changes critical records.

𝗪𝗛𝗔𝗧 𝗧𝗢 𝗪𝗔𝗧𝗖𝗛 𝗡𝗘𝗫𝗧

• Confirmed: OpenAI says Astra will expand from the initial limited rollout to paid ChatGPT plans and API customers over the coming days. Watch actual production reliability, usage allowances, and how often safety checks interrupt legitimate work.

• Confirmed: NVIDIA says Hugging Face will remain open and hardware-neutral after the proposed acquisition. Watch whether hosting terms, default infrastructure, model discovery, or multi-cloud support materially change after closing.

• Confirmed: The Anthropic/Pentagon dispute remains split across different legal authorities. Watch the D.C. Circuit case and any Defense Department procurement guidance that clarifies exactly where Claude is and is not restricted.