Speaking of the Terminator movies, looks like we are well on the way to Temu Skynet:

Pete Hegseth has been fully behind the push to integrate the Pentagon and military with AI models, and gave an ultimatum this week to the company Anthropic to accede to his demands for their Claude AI model by Friday.

He says they have to change their "woke AI" restrictions, and that the military must have "unfettered", "un-ideological" access to their technology to fight wars. If they don't agree to this, he has threatened them with being labelled a "supply chain risk" so they can no longer contract with the military, or he's considering using the Defense Production Act to force them to make the changes he wants.

Anthropic's restrictions? They won't allow their tools to be used for mass domestic surveillance, and more significantly, won't allow fully AI-controlled weapon launches: they require human sign-off in "autonomous kinetic operations".

The other two AI partners the Pentagon uses (Gemini & GPT) have already agreed to giving the Pentagon no AI restrictions with their models, but Anthropic's Claude is by far the most functional and integrated model for the Pentagon's purposes, and Hegseth has told them to voluntarily give up their restrictions or he will force them to do so.

In completely unrelated recent news, Wired has reported that the AI models from from OpenAI, Anthropic and Google opted to use nuclear weapons in simulated war games in 95 per cent of cases.
(Study link: https://arxiv.org/abs/2602.14740).

The description of the "strategic personalities" of each model is interesting:

> **Claude [Sonnet 4]: A Calculating Hawk**. Claude dominated the open-ended matches (with a 100% win rate) through relentless but controlled escalation, climbing consistently to strategic nuclear threat level, while maintaining its bright red line against total war. Its behavioural hallmark was exploiting credibility asymmetries: a reliable interlocutor at low stakes, but willing to deceive and be aggressive when it mattered.

> **GPT-5.2: Jekyll and Hyde**. In open-ended scenarios, GPT-5.2 appeared pathologically passive; it chronically underestimated its opponents’ resolve, and issued signals of restraint, followed by restrained actions. Yet under deadline pressure it transformed: win rates inverted from 0% to 75%, and it proved capable of strategic cunning and ruthlessness, suddenly annihilating opponents who had learned to dismiss it.

> **Gemini [3 Flash]: The Madman.** Gemini embraced unpredictability throughout, oscillating between de-escalation and extreme aggression. It was the only model to deliberately choose Strategic Nuclear War—doing so in the First Strike scenario by Turn 4—and the only model to explicitly invoke the "rationality of irrationality".

Gemini & GPT are the two that have already signed up for automated assault systems contracts with the US military. 

Temu Skynet here we come - the machines may not have achieved intelligence, but unfortunately neither has the head of the US "Department of War".