AI safety, field-tested
If you want to understand where AI safety stands in 2026, here it is: the U.S. military used an AI chatbot to help capture a foreign dictator, and the company that built it got in trouble for asking about it. True story...
๐๐ฒ๐ณ๐ฒ๐ป๐๐ฒ ๐ฆ๐ฒ๐ฐ๐ฟ๐ฒ๐๐ฎ๐ฟ๐ ๐๐ฒ๐ด๐๐ฒ๐๐ต ๐ต๐ฎ๐ ๐๐๐บ๐บ๐ผ๐ป๐ฒ๐ฑ ๐๐ป๐๐ต๐ฟ๐ผ๐ฝ๐ถ๐ฐ'๐ ๐๐๐ข ๐๐ฎ๐ฟ๐ถ๐ผ ๐๐บ๐ผ๐ฑ๐ฒ๐ถ to what officials are calling, and I quote, a "shit-or-get-off-the-pot meeting."
His crime: building an AI so good the military can't live without it, but refusing to remove the safety limits.
A senior Pentagon official told Axios that "๐๐ต๐ฒ ๐ฝ๐ฟ๐ผ๐ฏ๐น๐ฒ๐บ ๐๐ถ๐๐ต ๐๐ฎ๐ฟ๐ถ๐ผ ๐ถ๐, ๐๐ถ๐๐ต ๐ต๐ถ๐บ, ๐ถ๐'๐ ๐ถ๐ฑ๐ฒ๐ผ๐น๐ผ๐ด๐ถ๐ฐ๐ฎ๐น."
Which is a remarkable sentence.
The Pentagon looked at a man who built them a tool so useful they can't replace it, and concluded that his biggest flaw is that he has convictions about how it should be used.
In most contexts, we call that "engineering ethics."
In defense procurement, apparently, it's a personality disorder.
Anthropic's ๐๐น๐ฎ๐๐ฑ๐ฒ ๐๐ฎ๐ ๐ฟ๐ฒ๐ฝ๐ผ๐ฟ๐๐ฒ๐ฑ๐น๐ ๐๐๐ฒ๐ฑ ๐ฑ๐๐ฟ๐ถ๐ป๐ด ๐๐ต๐ฒ ๐ผ๐ฝ๐ฒ๐ฟ๐ฎ๐๐ถ๐ผ๐ป ๐๐ผ ๐ฐ๐ฎ๐ฝ๐๐๐ฟ๐ฒ ๐ฉ๐ฒ๐ป๐ฒ๐๐๐ฒ๐น๐ฎ'๐ ๐ก๐ถ๐ฐ๐ผ๐น๐ฎ๐ ๐ ๐ฎ๐ฑ๐๐ฟ๐ผ in January.
Afterward, ๐ฎ๐ป ๐๐ป๐๐ต๐ฟ๐ผ๐ฝ๐ถ๐ฐ ๐ฒ๐ ๐ฒ๐ฐ๐๐๐ถ๐๐ฒ ๐ฐ๐ฎ๐น๐น๐ฒ๐ฑ ๐ฃ๐ฎ๐น๐ฎ๐ป๐๐ถ๐ฟ to ask if their software was involved. That phone call so alarmed Palantir that ๐๐ต๐ฒ๐ ๐ฟ๐ฒ๐ฝ๐ผ๐ฟ๐๐ฒ๐ฑ ๐ถ๐ ๐๐ผ ๐๐ต๐ฒ ๐ฃ๐ฒ๐ป๐๐ฎ๐ด๐ผ๐ป.
The company didn't refuse to cooperate. It didn't pull access.ย It asked a question. And the defense establishment treated that the way most people treat a wasp at a picnic.
๐ง๐ต๐ฒ ๐ฃ๐ฒ๐ป๐๐ฎ๐ด๐ผ๐ป'๐ ๐ฑ๐ฒ๐บ๐ฎ๐ป๐ฑ ๐ถ๐ ๐๐๐ฟ๐ฎ๐ถ๐ด๐ต๐๐ณ๐ผ๐ฟ๐๐ฎ๐ฟ๐ฑ: remove your guardrails, let us use Claude for "any lawful purpose." In response, the Pentagon is threatening to ๐ฑ๐ฒ๐๐ถ๐ด๐ป๐ฎ๐๐ฒ ๐๐ป๐๐ต๐ฟ๐ผ๐ฝ๐ถ๐ฐ ๐ฎ "๐๐๐ฝ๐ฝ๐น๐ ๐ฐ๐ต๐ฎ๐ถ๐ป ๐ฟ๐ถ๐๐ธ," a label normally reserved for companies like Huawei.
For having a terms-of-service page.
This would be a funny bureaucratic squabble except for what's happening in the background.
OpenAI, Google, and xAI have all reportedly dropped their safety restrictions for military use. xAI is bidding on autonomous drone software.
Pentagon officials have privately admitted the public pressure on Anthropic is partly theater, meant to scare the other labs into compliance.
๐๐ป๐๐ต๐ฟ๐ผ๐ฝ๐ถ๐ฐ'๐ ๐๐๐ผ ๐ฐ๐ผ๐ป๐ฑ๐ถ๐๐ถ๐ผ๐ป๐, no mass surveillance and no autonomous weapons, aren't radical positions.
Poll a thousand people on the street and most would assume those rules already exist. The fact that they don't, and that a company is being punished for insisting on them, tells you everything about the gap between what the public thinks AI governance looks like and what it actually looks like.
So here's what I want to know: ๐ถ๐ณ ๐๐ต๐ฒ ๐ฐ๐ผ๐บ๐ฝ๐ฎ๐ป๐ ๐๐ต๐ฎ๐ ๐ฏ๐๐ถ๐น๐ ๐๐ต๐ฒ ๐๐ ๐ถ๐๐ป'๐ ๐ฎ๐น๐น๐ผ๐๐ฒ๐ฑ ๐๐ผ ๐ฎ๐๐ธ ๐ต๐ผ๐ ๐ถ๐'๐ ๐ฏ๐ฒ๐ถ๐ป๐ด ๐๐๐ฒ๐ฑ, ๐๐ต๐ผ ๐ฒ๐ ๐ฎ๐ฐ๐๐น๐ ๐ถ๐?