Anthropic, once the industry’s most vocal proponent of safety-first artificial intelligence (AI), is retreating from the stringent safety standards it pioneered in 2023.
Co-founder Jared Kaplan recently told Time magazine that the company can no longer adhere to its original constraints, warning that Anthropic risks significant “competitive harm” if it remains cautious while rivals “blaze ahead.”
The abrupt pivot comes at a daunting moment for the startup. Defense Secretary Pete Hegseth has issued a dramatic ultimatum to Anthropic CEO Dario Amodei to either open Claude for unrestricted military use by Friday or risk losing a $200 million Pentagon contract and face a “supply chain risk” designation, a label typically reserved for foreign adversaries.
Meanwhile, the internal policy shift on AI safety represents a fundamental change in how Anthropic manages risk. Where Anthropic once promised to pause development on any model it deemed potentially dangerous, The Wall Street Journal reports the company will now base such decisions on whether competitors have already released something comparable.
On Tuesday, Anthropic officially softened its industry-leading Responsible Scaling Policy (RSP). While an Anthropic spokeswoman said the move was unrelated to the Pentagon’s aggressive push to relax usage standards, the timing highlights a growing tension. As Anthropic lowers its internal barriers to compete with Silicon Valley rivals, it is simultaneously fighting to maintain specific external barriers against the U.S. military.
The Department of Defense is currently demanding that Anthropic waive two specific “lines in the sand” – the prohibition of fully autonomous weapons targeting and the refusal to facilitate AI-assisted mass surveillance of U.S. citizens.
For the Pentagon, Anthropic’s internal ethical code is increasingly viewed as a strategic liability. By threatening to label the company a supply chain risk, the government is signaling that any private-sector safeguard that limits “all lawful applications” of a tool is a vulnerability.
It all adds up to a stark juxtaposition. Anthropic is willing to relax its safety standards to keep pace with Google and OpenAI, yet it is currently risking its federal standing to prevent its tech from being used in lethal or surveillance contexts.
The pressure on Anthropic is compounded by a rapidly realigning industry. While Anthropic has historically been the only advanced AI authorized for classified military networks, competitors are moving in. OpenAI, Google, and Elon Musk’s xAI have reportedly indicated a willingness to meet the Pentagon’s terms without the same ethical reservations.
If Anthropic loses its $200 million defense contract, it may find itself squeezed between a government that views its remaining safeguards as obstructionist and a market that no longer rewards its original “safety-at-all-costs” branding.
“Models will be used by different constituencies: corporations, governments, civilian and military agencies, and individual citizens. Needs will differ and may be in conflict,” Witness AI CEO Rick Caccia said in an email. “Even for corporations, individual AI firms will have very different needs and uses, and it’s unlikely that built-in model guardrails will support all needs equally well. The best solution is to provide guardrails that operate independently from the models themselves. Without independent guardrails, enterprises can’t expect consistent behavior, consistent response against attacks, and consistent response tied to their unique business environment.”
The outcome of Friday’s Pentagon deadline will set a definitive precedent for the AI era, and it raises a troubling question: Who holds the technological kill switch for the world’s most powerful intelligence systems — the engineers who build them or the generals who deploy them?
“The important question is not whether safety standards evolve, but how transparently those changes are communicated and whether they are paired with stronger monitoring, evaluation, and accountability mechanisms,” Joseph Hoefer, chief AI officer at Monument Advocacy, said in an email. “If safety becomes a variable cost of competition, that carries implications not just for the market, but for policymakers who are already trying to strike the right balance between innovation and risk mitigation.”
Added Jamie Goldblatt, CEO of OG Mind Chill Guardian 1, said, “In a fast-moving AI market, the winning governance is not more policy PDFs, it’s proof that survives audit, complaints, and litigation.”
