In the span of five days, Anthropic—the company that built its entire brand on AI safety—did two things that seem completely contradictory.

First, it rewrote its Responsible Scaling Policy, dropping the hard commitment to pause AI development if safety measures can't keep up. The pledge that made Anthropic Anthropic is gone.

Then, it refused the Pentagon's demand to strip safety guardrails from its Claude model for military use. Specifically, two non-negotiables: no mass surveillance of Americans, and no fully autonomous weapons without a human in the loop. The Pentagon gave them until 5pm Friday. Anthropic said no. Trump ordered every federal agency to stop using Anthropic's technology. Hegseth designated the company a "supply chain risk"—a label normally reserved for foreign adversaries like Huawei.

Hours later, OpenAI struck a deal with the Pentagon that included the exact same red lines Anthropic was just punished for.

Bend on one thing. Break on another. The question for every business deploying AI right now is: do you know which is which?

Since 2023, Anthropic has operated under a system of AI Safety Levels. The idea was straightforward: as their models got more capable, harder safeguards would kick in. And if the models crossed a threshold where safety measures couldn't keep pace, the company would stop scaling. Period.

That commitment is gone in version 3.0.

It's worth understanding why before passing judgment. Anthropic spent nearly a year debating this internally. The original framework was built in September 2023 when large language models were mostly chat interfaces. In 2026, they're running classified military operations, writing production code, and managing autonomous business workflows. The landscape changed. The policy didn't fit anymore.

Three forces made the original design untenable. First, a political environment that has gone permissive on AI—no federal AI legislation anywhere on the horizon. Second, the science of evaluating AI risk is still ambiguous enough that it's hard to make a clean public case for when a model is "dangerous enough" to pause. And third, the safety measures required at higher levels would be essentially impossible for any single company to implement without industry-wide coordination. A RAND report cited in the policy describes its highest security standard as "currently not possible" to achieve without help from the national security community.

And there's the competitive pressure. Anthropic just raised $30 billion at roughly $380 billion valuation. Revenue is growing tenfold year over year. And competitors aren't operating under the same constraints. As the new policy states: if one developer pauses while others plow ahead without equivalent protections, "responsible developers would lose their ability to do safety research." It's not a comfortable argument, but it's a legitimate one.

The hard stop was replaced with public transparency. Anthropic now publishes a Frontier Safety Roadmap with concrete but nonbinding safety goals. They'll issue Risk Reports every three to six months with third-party expert review. And they've split the policy into two tracks: what Anthropic will do unilaterally, and what they think the industry as a whole needs to do.

Chris Painter at METR, an AI evaluation nonprofit, told TIME the change suggests Anthropic has shifted into "triage mode"—methods to assess risk aren't keeping pace with capabilities.

He's probably right. But a rigid rule that can't survive contact with reality isn't much of a safety measure. Anthropic adapted. Whether the replacement is good enough is a fair question. That they adapted at all is the right instinct.

The same week it loosened its development policy, Anthropic refused to budge when the U.S. military demanded it remove its remaining guardrails.

Anthropic exists because of AI safety. Dario Amodei was Vice President of Research at OpenAI. His sister Daniela was VP of Safety & Policy. In 2021, they walked out—along with five other senior researchers—because they believed OpenAI was scaling its models without adequate safety measures. As Dario put it at the time: the group believed "you needed something in addition to just scaling the models up, which is alignment or safety." They didn't write a memo. They didn't file a complaint. They left and built a competitor. That's the founding DNA of this company. So when the Pentagon showed up demanding Anthropic remove its guardrails, they weren't asking a company to change a policy. They were asking it to abandon its reason for existing.

Anthropic's Claude has been the only AI model deployed on the Pentagon's classified networks, integrated through a partnership with Palantir. The model was reportedly used in the operation to capture Venezuela's Nicolás Maduro.

The Pentagon wanted a blanket clause: Claude can be used for "all lawful purposes." Anthropic wanted two exceptions carved out. Just two: no mass domestic surveillance and no fully autonomous weapons without a human in the loop.

Defense Secretary Pete Hegseth summoned Anthropic's leadership to a meeting earlier in the week and set a Friday 5:01pm deadline. Agree to our terms or face consequences.

Anthropic held its position. CEO Dario Amodei published a statement saying the company "cannot in good conscience accede to their request." He pointed out the contradiction in the Pentagon's threats: labeling Anthropic a security risk and threatening to invoke the Defense Production Act to force use of Claude. As Amodei put it: "One labels us a security risk; the other labels Claude as essential to national security."

What followed escalated fast.

Friday afternoon: Trump posted on Truth Social: "I am directing EVERY Federal Agency in the United States Government to IMMEDIATELY CEASE all use of Anthropic's technology."

Minutes later: Hegseth designated Anthropic a supply chain risk, declaring that no military contractor, supplier, or partner "may conduct any commercial activity with Anthropic."

That evening: Anthropic responded that it had not yet received direct communication from either the Pentagon or the White House. It called the designation "legally unsound," said it would challenge it in court, and pointed out that Hegseth likely doesn't have statutory authority to extend the ban beyond DoD contracts.

Late Friday night: OpenAI CEO Sam Altman announced a deal with the Pentagon for classified deployment. His agreement includes the same red lines Anthropic was blacklisted for—plus a third restriction against high-stakes automated decisions like "social credit" systems. OpenAI keeps its own safety stack, deploys its own cleared personnel, and retains the right to strengthen its monitoring systems over time.

One detail worth sitting with: an Axios source revealed that the Pentagon's undersecretary Emil Michael was on the phone offering Anthropic a last-minute deal at the exact moment Hegseth tweeted the supply chain risk designation. That deal reportedly would have required Anthropic to allow the collection or analysis of Americans' geolocation data, web browsing history, and personal financial information purchased from data brokers.

Meanwhile, roughly 70 OpenAI employees signed an open letter in solidarity with Anthropic called "We Will Not Be Divided." Hundreds more from Google signed similar petitions. Altman himself told CNBC: "For all the differences I have with Anthropic, I mostly trust them as a company, and I think they really do care about safety."

Then something interesting happened. Over the weekend, Claude hit #1 on Apple's App Store, overtaking ChatGPT for the first time. Users launched a "Cancel ChatGPT" campaign across Reddit and X after the OpenAI deal. Standing on principle cost Anthropic a $200 million government contract—but it may be earning them something harder to buy: consumer trust.

OK. So you’re not Anthropic or OpenAI. You're not building foundation models or negotiating classified contracts with the Pentagon. But the underlying dynamics are exactly what you're going to face—at a different scale—as AI tools become more embedded in your operations.

The two stories are a case study in governance. One is about knowing when your rules need to evolve. The other is about knowing which rules don't bend, period.

The RSP lesson is about pragmatism. Anthropic built a safety policy that worked for 2023. By 2026, the world changed enough that the policy was becoming an obstacle rather than a safeguard. They adapted. They traded rigid rules for transparency and public accountability.

Your business is going to face the same thing. The AI governance framework you set up this year will probably need to change next year. The agents you deploy, the data access policies you write, the approval workflows you design—they'll need to flex as the technology evolves and as you learn what actually works in practice. That's not weakness. That's operational maturity.

The Pentagon lesson is about red lines. Anthropic was clear on two things. First, current AI models aren't reliable enough to make kill decisions without a human in the loop—that's a technical reality, not an opinion. Second, enabling mass surveillance conflicts with the values the company was built on—and that's a line no contract can justify crossing.

What are your non-negotiables? When you're deploying AI agents across your organization, you're going to be asked versions of the same question.

Can the AI agent access customer PII? What about financial data? Employee records? Does it get read access only, or can it take action? Who approves escalations? What data can leave your network?

There will be pressure to "just open it up." The agent works better with more data access. The vendor says their security is fine. The team wants to move faster. And for a lot of those decisions, the right call is to loosen up, test, iterate—the RSP approach. Evolve your guardrails as you learn.

But some of those decisions are red lines. Customer data that cannot be fed into third-party models. Financial records that cannot leave your environment. Automated actions that need a human sign-off before they execute. When you hit those, you need to be Anthropic-at-the-Pentagon, not Anthropic-on-the-RSP.

The companies that are going to get this right aren't the ones with the most aggressive AI deployments or the tightest lockdowns. They're the ones that can tell the difference between a guardrail that needs updating and a guardrail that exists for a reason.

Our sponsor this week is Deel. If you are a business expanding in multiple countries - you need to check them out. They can simplify your life!

This guide from Deel breaks down how to build one global hiring system. You’ll learn about assessment frameworks that scale, how to do headcount planning across regions, and even intake processes that work everywhere. As HR pros know, hiring in one country is hard enough. So let this free global hiring guide give you the tools you need to avoid global hiring headaches.

Download the free guide today

Anthropic showed us both sides of governance. They loosened their development commitments because the world changed and rigid rules were getting in the way. Then they held their ground against the most powerful institution on the planet because some lines exist for reasons that don't change—the same reasons the Amodei siblings walked out of OpenAI five years ago.

Both decisions are going to cost them. Dropping the hard pause commitment has already eroded trust in the safety community. Refusing the Pentagon just cost them a $200 million contract, a government blacklisting, and the headache of a legal battle they didn't want.

Every organization deploying AI is going to face versions of these decisions. The pressure to move fast will push you to loosen things that shouldn't be loosened. The fear of falling behind will tempt you to lock things down that need room to evolve.

The framework isn't complicated: know which rules are experiments and which are principles. Update the experiments. Defend the principles.

P.S: hit reply and let me know what hard lines is your company taking with AI and responsible use of AI?

Keep reading and learning and, LEAD the AI Revolution 💪

Hashi & The Context Window Team!