Anthropic Just Dropped Its Safety Pledge. Here's What It Means for Your Business.
The Company We Build On Just Changed the Rules
Full transparency: we use Claude every day. It powers our content generation, our analysis workflows, our development tools. When we recommend AI to clients, Claude is usually the first name out of our mouths.
So when TIME reported that Anthropic, the company behind Claude, scrapped the central promise of its flagship safety policy, we paid attention. Then the story got bigger. Much bigger.
Here's everything that unfolded this week, what it means, and what you should do about it.
What Anthropic Promised (and Just Took Back)
In 2023, Anthropic introduced something called the Responsible Scaling Policy (RSP). The core pledge was straightforward: Anthropic would never train an AI system unless it could guarantee in advance that adequate safety measures were in place.
That was the whole point. A hard line. A tripwire. If safety couldn't be demonstrated before training a more powerful model, training wouldn't happen.
As of this week, that promise is gone.
The new RSP v3.0, approved unanimously by CEO Dario Amodei and the board, replaces that hard commitment with something softer. Anthropic will now "delay" development only if leadership considers the company to be leading the AI race and believes catastrophic risks are significant.
That's a lot of conditions on what used to be unconditional.
Why They Did It
Anthropic's chief science officer Jared Kaplan put it bluntly to TIME: "We felt that it wouldn't actually help anyone for us to stop training AI models." He added that making unilateral safety commitments didn't make sense "if competitors are blazing ahead."
Three forces drove the change:
The race intensified. OpenAI, Google DeepMind, and others are pushing capabilities forward fast. Anthropic's original hope was that the RSP would inspire rivals to adopt similar measures. Some did, but nobody matched Anthropic's hard commitment to halt development. Pausing while others sprint isn't a safety strategy. It's a way to lose relevance.
Regulation never came. When the RSP launched in 2023, there was real momentum toward federal AI regulation and even international treaties. That momentum died. The Trump administration has endorsed what critics call a "let it rip" approach to AI development, even attempting to nullify state-level regulations. No federal AI law is on the horizon.
The science got harder. AI safety evaluations turned out to be more complicated than anyone expected. In 2025, Anthropic announced it couldn't rule out the possibility of its models facilitating bioterrorism, but also couldn't prove they did. What they imagined would be a bright red line turned out to be, in Kaplan's words, "a fuzzy gradient."
And then there's the Pentagon. During this same week, a separate but related conflict was escalating. Defense Secretary Pete Hegseth gave Anthropic until Friday to grant the military unrestricted access to Claude or face consequences, including invoking the Defense Production Act and canceling a $200 million contract. Claude had been the Pentagon's primary AI system for classified defense work. An Anthropic spokesperson told the Wall Street Journal the RSP update was unrelated to the Pentagon dispute, but the timing made it impossible to separate the two stories. (More on how this played out below.)
What the New Policy Actually Says
Let's be fair: Anthropic didn't abandon safety entirely. The new RSP v3.0 includes real commitments:
- Risk Reports published every 3-6 months, detailing how capabilities, threats, and mitigations fit together
- Frontier Safety Roadmaps with public goals across security, alignment, safeguards, and policy
- Competitive parity, a commitment to match or exceed the safety efforts of any competitor
- Conditional delays if leadership believes Anthropic is leading the race and catastrophic risks are significant
That's more transparency than most AI companies offer. But it's also fundamentally different from a hard commitment to stop.
The old policy said: We won't build it until we can prove it's safe.
The new policy says: We'll try really hard to make it safe, and we'll tell you about it.
Those aren't the same thing.
The "Frog-Boiling" Problem
Chris Painter, director of policy at METR (a nonprofit focused on evaluating AI models for risky behavior), reviewed the new policy. His reaction was measured but sobering.
He warned that removing binary thresholds, the hard tripwires that would halt development, could enable a "frog-boiling" effect, where danger ramps up gradually without any single moment that sets off alarms.
His assessment of what the change signals: "This is more evidence that society is not prepared for the potential catastrophic risks posed by AI."
Painter also noted that the shift shows Anthropic "believes it needs to shift into triage mode with its safety plans, because methods to assess and mitigate risk are not keeping up with the pace of capabilities."
That last part is the most important. The company building some of the most capable AI in the world is telling us, publicly, that safety research can't keep up with capability research.
Then Friday Happened
The RSP story would have been significant on its own. Then the Pentagon deadline hit, and this became the biggest story in AI.
On Wednesday, CEO Dario Amodei published a statement making Anthropic's position unambiguous: the company "cannot in good conscience accede" to the Pentagon's demand for unrestricted military access to Claude. Specifically, Anthropic refused to allow its AI to power fully autonomous weapons systems or mass domestic surveillance of Americans.
He called the Pentagon's threats "inherently contradictory: one labels us a security risk; the other labels Claude as essential to national security." He added: "Regardless, these threats do not change our position."
On Friday, the consequences landed. President Trump ordered every federal agency to "immediately cease" using Anthropic's technology. Hegseth designated Anthropic a "supply chain risk to national security," a move that would bar any military contractor from doing business with the company. Federal agencies have six months to phase out all Anthropic products.
Hours later, OpenAI announced a deal to deploy its models on the Pentagon's classified networks. CEO Sam Altman said the Department of Defense agreed to restrictions on "domestic mass surveillance and human responsibility for the use of force, including for autonomous weapon systems."
Pause on that. The Pentagon punished Anthropic for demanding guardrails on surveillance and autonomous weapons, then turned around and agreed to those same guardrails with OpenAI.
Anthropic has announced it will challenge the supply chain risk designation in court, calling it "legally unsound" and "a dangerous precedent for any American company that negotiates with the government."
Two Kinds of "Safety"
Here's the nuance most coverage is missing: the RSP change and the Pentagon standoff involve two different kinds of safety.
The RSP governs development safety: under what conditions will Anthropic train more powerful models? On this front, Anthropic loosened its stance. The hard commitment to halt training without proven safety measures is gone.
The Pentagon dispute is about use restrictions: what can people do with Claude once it exists? On this front, Anthropic held firm. They refused to remove prohibitions on autonomous weapons and mass surveillance, and they're paying a steep price for it.
One commitment weakened. Another held. The full picture is more complicated than either "Anthropic caved on safety" or "Anthropic stood up for safety."
What This Means If You're Using AI in Your Business
Let's bring this down to earth. If you're a business owner using Claude (or any AI tool), here's our honest take:
Your tools aren't suddenly less safe today. Claude didn't change overnight. The models you're using right now are the same ones you were using yesterday. This is a policy change about future development, not a product recall.
But the safety net just got thinner. The guarantee that Anthropic would hit the brakes if safety couldn't be demonstrated? That's gone. What remains is a promise to try hard and be transparent about it. That's better than nothing, but it's not the same as a hard stop.
This isn't just about Anthropic. Within months of Anthropic's original RSP in 2023, both OpenAI and Google DeepMind adopted broadly similar frameworks. A rollback by the policy's originator may reshape what "responsible scaling" means across the entire industry. If the company that set the standard loosens it, everyone else will feel permission to do the same.
Platform risk just became real. If you're a federal contractor or government-adjacent business running on Claude, this week was a wake-up call. One executive order can cut off your AI provider overnight. Even in the private sector, this is a reminder that building your entire operation on a single AI vendor creates concentration risk. Have contingency plans. Know what your backup looks like.
You need to be your own safety layer. This has always been true, but it matters more now. Don't outsource your judgment to any AI company's safety promises. Build review processes into your workflows. Keep humans in the loop for anything that touches customers, finances, or critical decisions.
We wrote about this in Why 95% of AI Projects Fail, the businesses that succeed with AI are the ones that build for failure. They assume AI will make mistakes and design systems that catch them before they cause damage.
That principle just got more important.
Our Position
We're not abandoning Claude. We're not telling you to either. It remains, in our experience, the most capable and thoughtful AI platform available.
This week actually reinforced that in unexpected ways. Anthropic loosened its scaling policy, and that concerned us. But when the Pentagon demanded the right to use Claude for autonomous weapons and mass surveillance, Anthropic said no and accepted the consequences. That took real conviction, especially when their biggest competitor immediately stepped in to fill the gap.
We're going into this next chapter with eyes wide open. The rules are changing fast. Safety commitments are softening. Government pressure is intensifying. And the companies building the most powerful AI in the world are being forced to choose between principles and market position.
That means the responsibility shifts. To us. To you. To every business building on these tools.
Build your own guardrails. Keep humans where judgment matters. Diversify your AI stack. Don't let any single vendor or government decision determine your ability to operate.
The AI tools are getting more powerful. The forces shaping how they're used are getting more complicated. The question is whether we're keeping up.
Want to make sure your AI workflows have proper human oversight built in? Book a free 30-minute call and let's talk about building AI processes that work, safely.
Sources:
- TIME: Anthropic Drops Flagship Safety Pledge (February 25, 2026)
- Anthropic: Responsible Scaling Policy v3.0 (February 24, 2026)
- Bloomberg: Anthropic Drops Hallmark Safety Pledge in Race With AI Peers (February 25, 2026)
- CNN: Anthropic ditches its core safety promise (February 25, 2026)
- Engadget: Anthropic weakens safety pledge in wake of Pentagon pressure (February 25, 2026)
- Fortune: Dario Amodei says he "cannot in good conscience" bow to Pentagon's demands (February 27, 2026)
- Fortune: Trump orders U.S. government to stop using Anthropic (February 27, 2026)
- NPR: OpenAI announces Pentagon deal after Trump bans Anthropic (February 27, 2026)
- CNBC: Trump blacklists Anthropic as AI firm refuses Pentagon demands (February 27, 2026)