One court ruling just redrew the boundary between government power and an AI company’s right to say no. That is the sharpest development in AI news today: a federal judge ruled that the Pentagon’s effort to blacklist Anthropic was unlawful retaliation, giving the Claude maker a decisive legal win after months of conflict over military use. On the same U.S. news day, Anthropic pushed agents beyond screens and into laboratories, while more than 100 organizations warned that increasingly capable models will also make cyberattacks easier. If you build, buy, govern, or invest in AI, the theme is not “smarter models.” It is control: who sets the limits, who holds the credentials, and who is accountable when software can act in the world. Coffee helps. A permission model helps more.

AI news today: Anthropic wins a consequential Pentagon case

U.S. District Judge Rita F. Lin ruled on August 27 that the Pentagon’s supply-chain-risk designation for Anthropic was contrary to law and arbitrary and capricious. Her 59-page order found that the government retaliated against Anthropic’s protected speech after the company opposed unrestricted use of Claude for mass domestic surveillance and fully autonomous weapons. The court granted Anthropic summary judgment on key First Amendment, due-process, and Administrative Procedure Act claims, while granting the government judgment on some other claims.

The direct answer is clear: the government may choose another AI supplier, but it cannot brand a company a national-security risk as punishment for publicly disagreeing with it. An appeal remains possible. Still, the ruling strengthens model providers that want contractual red lines. For leaders, usage policies are becoming part of product architecture, procurement, and constitutional litigation—not a page nobody reads until something catches fire. Our earlier analysis of agentic AI governance explains why those controls now sit at board level.

116 organizations call for a cyber-defense surge

OpenAI, Anthropic, Google, Microsoft, Amazon Web Services, Cloudflare, CrowdStrike, banks, manufacturers, and public-interest groups joined an open letter calling for collective action on cyber defense. The signatories say there is a limited window to harden critical systems before AI-enabled attacks become more widespread. Their proposals include giving vetted defenders trusted access to cyber-capable models, sharing verified fixes, prioritizing defense at leadership level, and building cross-sector partnerships.

This is a forecast and policy appeal, not proof that a specific wave of attacks has arrived. Still, the coalition’s breadth matters. The awkward tension is that several signatories are also building the capabilities behind the warning. Watch whether the commitments produce measurable funding, access programs, and improvements for hospitals, water utilities, and smaller organizations—not merely another handsome PDF with excellent logos.

Anthropic gives AI agents a path into physical laboratories

Anthropic opened a research preview of the Model Hardware Standard, or MHS, a model-agnostic specification for agents to operate programmable equipment. Anthropic says participating labs and manufacturers can connect microscopes, liquid handlers, robotic arms, and quantum-computing hardware in hours or minutes instead of weeks or months. Agents can coordinate instruments, adjust parameters, and sometimes recover from hardware errors.

Those are company claims from an early preview, not independently established guarantees. Access is limited, and Anthropic says it will work with partners on safety evaluations before open-sourcing the standard. Even so, MHS extends the interoperability logic of Model Context Protocol into machines that can spill, break, heat, cut, or move. Builders should watch permissions and fail-safe design at least as closely as automation demos. The transition from “agent made the wrong spreadsheet” to “agent moved the wrong robot arm” deserves a different risk register.

OpenAI is reportedly testing an agent that does not stop

WIRED reports that OpenAI is developing a “Persistent mode” for Codex that could keep working until a user stops it, generate follow-up tasks, and issue occasional notifications. This is reported testing, not a launched product; OpenAI has not published general availability, pricing, or final behavior.

The direction matters. Long-running agents shift the bottleneck from prompting to supervision: clear scopes, spend limits, audit logs, revocable credentials, and reliable stop controls. A useful agent can compound progress; a confused one can compound damage with equal enthusiasm. See our guide to OpenAI Codex and isolated agent work for context. What comes next is evidence that persistent operation improves completed work without making oversight impractical.

Workday says AI drives more than a quarter of new contract value

Workday reported fiscal second-quarter revenue of $2.649 billion, up 12.8% year over year, and said AI generated more than 25% of new annual contract value. More than 5,500 customers now use at least one Workday-built agent, up over 35% from the prior quarter, according to the company’s earnings release. Workday also highlighted Agent Passport, designed to test and continuously monitor first- and third-party agents.

These are management-reported figures, but they are more useful than vague engagement metrics because they connect agents to contracts. Enterprise AI is beginning to sell as governed workflow, not only as a chat box. Watch whether adoption produces durable subscription growth and cash flow—not just larger bundles and friendlier terminology.

Marvell’s data-center revenue jumps 46%

Marvell closed the day with another hard number for the AI infrastructure cycle. The chipmaker reported record quarterly revenue of $2.739 billion, up 37% year over year, while data-center revenue reached $2.172 billion, up 46%. It raised its fiscal 2027 and 2028 revenue outlooks and guided the next quarter to roughly $3.15 billion, according to Marvell’s results.

Marvell attributes the acceleration to connectivity and custom silicon demand. That supports the view that AI spending is spreading beyond GPUs into networking, optical links, and chips tailored for hyperscalers. Investors still need to separate demand from profitability: GAAP gross margin was 53.1%, versus 58.9% non-GAAP. The October 6 investor day should reveal how much of the custom-chip pipeline is contracted, how concentrated it is, and when it converts into revenue.

Watch & Learn

Editor’s note: Docker’s 51-minute tutorial, “How Docker Sandboxes AI Agents (Before They Break Everything),” explains how microVM and container isolation can limit what coding agents touch. It is best for developers and technical leaders evaluating long-running agents. You will leave with a practical model for separating convenience, credentials, files, and execution risk.

AI, Translated: sandboxing

Sandboxing means running software inside an isolated environment with deliberately limited access to the rest of a computer or network. For example, a coding agent can install packages and run tests in a temporary container without receiving your production database password or unrestricted laptop access. If the agent follows a malicious instruction from a web page, the sandbox narrows the damage. You should care because more capable and persistent agents need smaller blast radiuses, not larger circles of trust.

Try This Today: audit a connected workflow in Grok

Goal: find where an assistant could overreach before connecting it to work data. In Grok, open the Connectors page, choose a low-risk source such as a test Google Drive folder, and complete OAuth. Business and Enterprise workspaces require an admin to provision connectors first. Then:

1. Ask Grok to list what the connector can read or change.
2. Give it a read-only task and require citations to exact source files.
3. Ask which requested actions would need separate approval.

Copy-ready prompt: “Using only the connected test folder, summarize the three newest documents. Cite each file. Do not create, edit, move, delete, email, or share anything. Finish with a permission audit: what you accessed, what you did not access, and which next actions would require my approval.” This takes about ten minutes and turns “the connector works” into a small, inspectable security test.

One thing to remember

AI’s next phase is not defined only by what agents can do. It will be defined by who can stop them, what they may touch, and whether those limits survive contact with governments, customers, and the physical world.