Something went wrong after Operation Absolute Resolve. Or rather — something went right, the mission succeeded, and then somebody at Anthropic made a phone call that undid quite a lot of the goodwill.
An executive at the San Francisco company behind Claude contacted Palantir Technologies in the days after the Caracas raid to ask a question that sounds reasonable enough on paper: how had their AI model been used during a classified military operation that ended with people being shot? Palantir is the defence analytics firm through which Claude operates on Pentagon systems. The call, in other words, went to the middleman.
That call may cost Anthropic a $200m defence contract.
Two sources with direct knowledge of the situation told Axios the US military used Claude during the 3 January raid — making it, as far as anyone can confirm, the first time a commercially built AI model has been deployed inside a classified American military operation. The Wall Street Journal broke the story first, citing people familiar with the matter. What Claude actually did during the operation has not been made public; whether it crunched satellite imagery, ran logistics calculations, processed intelligence feeds or did something else entirely is unclear, and the classified nature of the mission means nobody involved has to explain.
Nobody has.
One Phone Call, Then the Knives Came Out
Here is where it gets properly messy. A senior administration official told Axios that the Anthropic executive’s call was ‘raised in such a way to imply that they might disapprove of their software being used, because obviously there was kinetic fire during that raid, people were shot.’
Anthropic flatly denied that version. A spokesperson insisted the company had ‘not discussed the use of Claude for specific operations with the Department of War.’
Nobody budged.
But the damage — or whatever the administration decided to call it, and they’ve been rather careful not to call it anything directly — was done. The Pentagon is now weighing whether to sever the relationship entirely: ‘Any company that would jeopardise the operational success of our warfighters in the field is one we need to reevaluate our partnership with going forward,’ the senior official told Axios.
Whether that official had been waiting weeks for a reason to say something like that is impossible to know from the outside. But the speed of the response suggests the frustration did not start with the phone call. It had been building.
The contract at stake was awarded last summer. Anthropic was the first AI model developer to operate on classified Pentagon systems, and the same official conceded it would be difficult to replace Claude quickly because ‘the other model companies are just behind’ on specialist government work. So the Pentagon is threatening to fire a contractor it cannot easily replace. The maths on that don’t quite add up, but since when has that stopped anyone?
What Anthropic Won’t Give Up
Strip away the Venezuela drama and the fight is about something specific: can a company sell its technology to the US government and still tell them not to use it for certain things? That is not a hypothetical any more.
The Pentagon wants all four major AI labs — Anthropic, OpenAI, Google and Elon Musk’s xAI — to make their models available on classified networks for ‘all lawful purposes,’ without the safety guardrails that apply to civilian users. OpenAI agreed. Google agreed. xAI agreed.
Anthropic did not.
Two red lines remain: no mass surveillance of American citizens, and no fully autonomous weapons systems. The company’s spokesperson told Axios that their negotiations ‘have focused on a specific set of Usage Policy questions — namely, our hard limits around fully autonomous weapons and mass domestic surveillance — none of which relate to current operations.’
A senior administration official described Anthropic as the most ‘ideological’ of the four labs. Coming from a government that had just deployed an AI chatbot during a military raid on a foreign capital, that word lands a bit oddly — though the official who used it probably didn’t see the irony. Mind you, they also conceded in the same breath that Anthropic is the hardest to replace. Draw your own conclusions from that.
The Raid
Before the phone call, before the contract dispute, before any of this became a story about procurement and principles — there was the operation itself.
It began around 2am local time on 3 January 2026. US Delta Force operators, backed by more than 150 aircraft, struck targets across the Venezuelan capital simultaneously. Air defence systems were knocked out within minutes. Electricity across Caracas went dark; US President Donald Trump later suggested the Americans had cut it deliberately, which is the kind of detail that gets dropped into a press conference and never quite confirmed or denied.
Helicopters carrying the apprehension team flew in at roughly 30 metres above the water. They reached Fort Tiuna — Maduro’s military compound on the southern edge of the city — and breached it in what Defence Secretary Pete Hegseth would later call a ‘massive joint military and law enforcement raid.’ The whole thing lasted under half an hour.
Maduro and his wife Cilia Flores were seized. Flown to the USS Iwo Jima. Transferred to New York. Both pleaded not guilty to narcoterrorism and drug trafficking charges at a Manhattan federal court on 5 January. Maduro told the judge he considered himself ‘a prisoner of war’ who had been ‘captured at my home in Caracas.’
Seven US service members were injured. No Americans died.
On the Venezuelan side, the figures depend on who you believe and when they said it. Defence minister Vladimir Padrino López put the death toll at 83 — including 47 Venezuelan troops and 32 Cuban soldiers assigned to Maduro’s personal protection unit. The Washington Post, citing US officials, said approximately 75. Interior minister Diosdado Cabello claimed more than 100, and said some of the dead were so badly damaged that DNA testing was required for identification. The New York Times initially reported 40 on the night of the raid, revised to 80 within 24 hours.
Whichever figure is closest to the truth, this was a lethal night in Caracas. And Claude — the same AI you can use to summarise a PDF or draft a polite email to your landlord — was running somewhere inside it.
What Happens Now
Hegseth laid out the Pentagon’s broader position back in December, when he launched GenAI.mil — a platform designed to put commercial AI on every military desktop in the country. ‘The future of American warfare is here, and it’s spelled AI,’ he said in a video statement on 9 December 2025. Google’s Gemini went live on the platform first. xAI’s Grok followed in January.
Anthropic’s Claude was already running on classified systems before either of those. That head start is now the problem; the company that got there first is the one being told to fall in line or get out. Anthropic CEO Dario Amodei has repeatedly warned about the dangers of autonomous lethal systems, and the company built its reputation on being the safety-conscious lab. But the Pentagon, it turns out, does not want a supplier that phones up after a raid to ask what happened.
Whether Anthropic can hold its two red lines and keep a $200m Pentagon contract is the question the company faces this month. Whether the Pentagon should want an AI partner that doesn’t ask uncomfortable questions after a classified operation — well, that’s a rather different question, and not one anybody in Washington seems especially keen to answer.
Discussions about the contract’s future are ongoing. No decision has been announced.
Originally published on IBTimes UK
