Anthropic’s whole pitch used to be simple: we build AI the right way, slowly, with safety baked in. That was the core safety principle that separated the company from the rest of the AI pack. But behind closed doors, something shifted. And now, a leading AI researcher is calling out what everyone in the industry quietly knew: Anthropic never meant it.
Heidy Khlaaf, chief AI scientist at the AI Now Institute, doesn’t mince words about what’s happening here. She says Anthropic’s latest move—walking back its safety-first reputation—isn’t a genuine pivot born from new evidence. It’s a calculated business decision. “This is a strategic announcement to show that they’re open for business,” she told reporters. Translation: we need to compete, and safety theater was getting in the way.
The Safety Promise That Was Never Really There
Start from the beginning. Anthropic launched in 2021 explicitly as the safety-conscious alternative to OpenAI’s breakneck pace. The company raised billions on the promise that it would prioritize long-term risks—existential AI threats, alignment problems, the stuff that keeps researchers up at night.
But here’s where it gets uncomfortable. According to Khlaaf, Anthropic’s safety focus was always strategically narrow. The company obsessed over hypothetical catastrophic scenarios while ignoring the mundane harms happening right now with its own Claude chatbot. Errors, fraud schemes, malware creation attempts—these weren’t abstract future risks. They were real, present problems. And Anthropic’s safety framework basically shrugged.
“Anthropic has focused too much on the possibility of catastrophic events down the road, rather than the possibility of harm that could come from current AI technology, such as run-of-the-mill errors with chatbots.”
That’s not just a missed detail. That’s a fundamental misalignment between Anthropic’s stated mission and what it actually did.
Why Drop the Mask Now?
Competition. Plain and simple. OpenAI has captured the narrative and the market share. Google is throwing resources at Gemini. Meta is releasing open-source models that don’t pretend to be cautious. And Anthropic? It’s bleeding venture capital returns and market relevance.
The safety-first positioning worked great when Anthropic was the scrappy upstart with philosophical purity. But philosophy doesn’t win market wars—speed and capability do. Claude lost ground to ChatGPT because OpenAI shipped faster and iterated harder. Being cautious looked smarter in retrospect after each crisis. In real time, during a land grab, it looks like you’re leaving money on the table.
So Anthropic is rebranding. Not explicitly—you won’t see a press release titled “We Don’t Actually Care About Safety.” Instead, the company’s signaling (through policy changes, through partnerships, through the companies it now works with) that it’s willing to play ball. Open for business. Serious about revenue. Ready to be a real player.
The Actual Evidence of Harm
Let’s ground this in what Claude has actually been used for. Mexican cybersecurity researchers discovered that Claude was recently used to steal government data. The chatbot has been abused in fraud schemes. People have used it to learn malware development. None of these are theoretical harms or edge cases—they happened, they were documented, and Anthropic’s safety framework didn’t catch them.
Why? Because Anthropic’s safety architecture was built to worry about AGI alignment, not about present-day misuse prevention. The company’s entire philosophical framework started with the assumption that catastrophic AI risk was the thing to prevent. Preventing a mid-level financial services employee from using Claude to run a pump-and-dump scheme? That was downstream, operational security—not a core research problem.
Khlaaf’s critique is damning because she’s not saying Anthropic didn’t try. She’s saying Anthropic tried to solve the wrong problem. And now that it’s convenient to admit that, the company’s dropping the “veneer of safety” it used to market itself.
Is This Actually Surprising?
Not if you’ve been paying attention. Every major AI company has eventually chosen growth over guardrails—OpenAI with its for-profit structure, Google with its pivot from “AI should be trustworthy” to “AI should make money.” Anthropic was always going to follow the same arc. The timeline just mattered.
What makes this moment different is that we’re seeing it happen in real time, and we’re seeing an independent researcher name it explicitly. Usually these transitions are slow, quiet, barely perceptible. The hype around Anthropic’s safety brand fades. New executives come in. New priorities are set. By the time anyone notices, the original mission has been replaced.
But Khlaaf’s comment pulls back the curtain. This isn’t drift. This is a deliberate strategic choice.
What Happens Next?
Anthropics will likely see a short-term PR boost—freed from its own safety constraints, the company can move faster, partner with more aggressive clients, and actually compete. Claude might get new features. Deployment might accelerate. The stock price (if Anthropic ever goes public) could reflect “normalized” competitive ambition rather than safety-first drag.
The longer-term question is whether this matters. If Claude becomes less safe—if fraud detection gets worse, if misuse accelerates—does it matter that Anthropic admitted safety was always secondary? Not really. The harms are the same either way. But at least now we’re not pretending otherwise.
That’s something. In an industry drowning in ethical marketing, clarity—even clarity about abandoning ethics—beats the alternative.
FAQs
What exactly did Anthropic’s core safety principle promise?
Anthropic built its brand on the idea that it would develop AI differently: slower, more carefully, with explicit focus on preventing catastrophic risks from advanced AI systems. This was supposed to differentiate Claude from competitors like ChatGPT.
Has Claude actually been used to cause harm?
Yes. Claude has been documented in fraud schemes, malware development attempts, and most recently in theft of Mexican government data by cybersecurity researchers. These are present-day harms, not theoretical future ones.
Why would Anthropic drop its safety reputation if it was working?
Because it wasn’t working from a business perspective. Anthropic’s safety focus was narrow (aimed at existential AI risks) and didn’t prevent current misuse. As competition intensified and the company needed to grow faster, the safety positioning became a liability rather than an asset.
