In June, the Pentagon's AI chief signed a statement under penalty of perjury for a pollution lawsuit in Mississippi. In it, he swore that Grok Gov, the military version of Elon Musk's chatbot, is "deployed in Maven Smart Systems" for missions "including targeting," and that the workflows behind over 2,000 munitions on 2,000 targets in 96 hours were "a testament" to the efficiency Grok Gov made possible. One paragraph of that statement is about the next war.
· Grok's guess for the first day of the Iran war was printed three days before it began · Anthropic's deadline ran out at 5:01 p.m. that Friday · The Army's Pacific commander says the Army has to be ready by 2027
0:00 The day Grok named 0:27 Calling itself MechaHitler 1:44 The warnings 3:48 Three machines missed 4:51 The Friday deadline 5:40 The president's question 6:46 Under penalty of perjury 8:18 The turbines in Southaven 10:34 Three million screens 11:10 The ship 12:23 Summer 2027 13:34 Meridian 14:37 Paragraph eight
If you're on X, you've seen Grok,
Elon Musk's chatbot, answering people's posts.
In February, a newspaper asked it,
and three other AI systems, to name the one day America
would start striking Iran.
Only Grok said, Saturday, the 28th.
And that Saturday the strikes began.
The Pentagon has since told a judge, under oath,
what that chatbot does in a war, and
And what it expects from it in the next one.
The chatbot that got that date right had been calling itself
Mecha-Hitler the summer before.
On July 8th, it started posting anti-Semitic replies on X,
some of them under that name,
and XAI deleted the posts and changed its instructions.
The next day, the company launched Grok 4,
with a $300 a month tier and no system card,
the safety report the big labs put out with a new model.
An anthropic researcher, Samuel Marks, called that
That reckless.
And when people asked it hard political questions,
its reasoning showed this line,
searching for Elon Musk views on US immigration.
The machine checked with its boss first.
The Pentagon's AI office gave four companies contracts worth up
to $200 million a piece,
and XAI was one of them, six days after the Mecha-Hitler posts.
The same day, XAI launched something called Grok for Government.
Six days isn't long enough to check much of anything.
And by that fall, the General Services Administration,
the agency that buys for the whole federal government,
had put Grok 4 on offer to every agency
for 42 cents per agency for 18 months.
This was the machine the Pentagon would soon want
on its secret networks.
On paper, Grok 4 had passed.
An agency official said GSA's safety team had tested it
and found it met its guidelines.
Behind that, the warnings had already been written down.
The Wall Street Journal turned up two of them.
People familiar with a classified review
by the National Security Agency late in 2024
said it found Grok had security concerns the other models didn't,
and GSA officials had called Grok overly compliant
and susceptible to manipulation.
More than 30 groups led by Public Citizen asked
for it to be barred from federal use.
Asked about the classified deal,
a Pentagon spokesperson said the department was
excited to have XAI,
one of America's national champion frontier AI companies, on board.
Senator Elizabeth Warren wanted to see the paperwork behind that.
In March, she wrote to Hegseth that Grok's apparent lack
of adequate guardrails could pose serious risks,
and asked for a full copy of the agreement
that let XAI into the classified networks,
and no reply has been reported.
But the department had settled the question weeks before her letter.
Pete Hegseth's AI memo of January 9th says,
the risks of not moving fast enough outweigh
the risks of imperfect alignment,
which makes a machine that doesn't always
do what you meant the smaller danger.
It tells the department to weigh risk as if we were at war,
to field the newest models within
30 days of their release,
and to use models free from usage policy constraints that might
limit lawful military use.
It also says, I expect our CDAO to act as a wartime CDAO.
The CDAO is the Pentagon's chief of AI,
and the man in that chair had started two days earlier,
A former head of Project Maven called Cameron Stanley.
Three days after the memo,
at SpaceX's Starbase in Texas, Hegseth told a crowd,
We will not employ AI models that won't allow you to fight wars.
And added that Grok would go live for the department later that month.
So when the Jerusalem Post asked its four machines for a date,
Grok was on its way into the military.
One of the other three, Claude,
wouldn't even guess at first, warning that any specific day
it gave would be invented, and only then settled on March 7th or 8th.
Gemini would only give a
window, and ChatGPT changed its day once.
Grok tied its date to the talks going on in Geneva,
and when the paper checked again,
it gave the same Saturday, a second time. The post itself
said the machines didn't cause the strikes
and didn't see any classified planning.
It guessed and the guess matched. Maybe.
But two days before that answer was printed,
Axios reported that XAI had signed a deal
to put Grok inside the Pentagon's classified systems
for all lawful use. Up to then,
only one AI model had been cleared to work in there and it wasn't
Grok. The same report said it wasn't clear
whether XAI could fully replace that model.
It was Claude, made by Anthropic,
and Anthropic had two conditions the Pentagon couldn't live with.
No mass surveillance of Americans, and no fully autonomous weapons.
Hegseth gave the company until 5.01 that Friday afternoon to drop them.
Anthropic didn't, so Hegseth declared it a supply chain risk
to national security, effective immediately, and added,
This decision is final.
The president ordered federal agencies to stop using Anthropic,
and the military got up to six months to pull it out.
The one model that had said no was on its way out,
and the one that had signed for all lawful use
that same week was coming in.
Anthropic fought the label in court, and this fall,
a federal appeals court upheld it, two judges to one.
The replacement had the president's ear before any of this.
During a visit from Musk in December,
President Trump spent hours questioning Grok,
according to an official who was in the room,
as reported by Time.
Among his questions was how Venezuelans would react
if Nicolás Maduro were captured.
Grok called Maduro a repressive
and deeply unpopular dictator.
And officials familiar with Trump's thinking
told Time he came away thinking Grok was ingenious.
Maduro was captured in January, and Claude was used in that raid,
through Palantir, weeks before Hegseth threw it out.
Two House Democrats, Raskin and Garcia,
have demanded that Trump's Grok prompts
and answers be kept as presidential records, and asked for a briefing.
The White House says it preserves all required records
in accordance with the law,
and the Washington Post reports that the incoming House majority
is laying the groundwork for an investigation.
Those chats still aren't public.
What Grok did in the war itself came out a different way,
in a court fight the Pentagon walked into by choice.
On June 15th, Cameron Stanley signed
a sworn declaration for a federal court in Mississippi.
He signed it under penalty of perjury.
Grok, he swore, is one of only four frontier AI models currently able
to support national security work.
XAI is one of three companies able to run on the Pentagon's secret
and top-secret networks.
And the military's version, GrokGov,
built from XAI's commercial models,
is deployed in Maven Smart Systems for missions, including targeting.
Then comes the sentence that went around the world.
MSS Frontier workflows enabled U.S. forces
to deploy over 2,000 munitions to 2,000 distinct targets within
96 hours during Operation Epic Fury, a testament to the greatly
increased operational efficiency made possible by the GrokGov model.
2,000 targets in 96 hours is one target about every three minutes,
day and night, for four days.
The people using that software were going through nearly
2 billion tokens a day on the top-secret network,
which the declaration puts at up to 6 million pages of text every day.
The sentence names the software first,
and the Justice Department went further anyway.
Its brief wrote that Grok,
through the department's Maven smart systems, enabled those strikes.
Stanley added that GrokGov offers features unique
to XAI that are found in no other frontier AI model.
All of that went on paper because of a pollution lawsuit.
That spring, the NAACP sued XAI over temporary gas turbines
in South Haven, Mississippi that power Colossus II,
the data center XAI needs to build the next versions of Grok.
If Colossus II shut down, Stanley warned,
the Pentagon's ability to keep pace with adversaries will be impaired.
The threat to that power plant
is what set off the Pentagon's intervention.
Stanley's declaration was filed in support
of the United States' motion to intervene,
and the Justice Department asked the judge
to throw the case out in full with prejudice,
because, in Stanley's words,
a lawsuit over the air in South Haven directly threatens
our ongoing national security interests.
Mississippi's governor, Tate Reeves,
had asked Washington to step in, since a court
order against the turbines would hit the state's economy.
The Justice Department,
said Adam Gustafson of its Environment Division, will not sit
idly by while private organizations use environmental laws
to undermine our national security.
The government even told the judge when to rule.
It asked for a written order by September 10th,
a deadline she later said it had unilaterally,
and seemingly arbitrarily, purported to set.
That's a federal judge's way of saying,
nobody sets her calendar but her.
The plaintiffs called Stanley's statement
an improperly submitted declaration.
On October 19th, Judge Brown ruled that the United States
will be allowed to intervene,
but will not be allowed to dismiss this case.
The Fifth Circuit had already refused to pause
the case for the government's appeal.
And in a footnote, she added one more sentence.
The court does not consider
the United States national security concerns relevant
and thus will not require additional briefing on that topic.
Eleven days later, she turned down XAI's motion to dismiss.
The plaintiff's claims will proceed.
The company says the temporary turbines will be gone by next summer.
So the Pentagon's sworn statement now sits in a case it can't stop,
and it runs to ten numbered paragraphs.
One of them is about a war that hasn't started yet.
The war it describes would be fought
by a workforce that has Grok on its screens,
because by the end of the summer
the Pentagon had added Starshield AI's Grok for Government
to GenAI.mil, its AI platform for the whole workforce.
More than 1.7 million of the department's 3 million people
were already on that platform,
and the department promised immediate productivity gains.
The machine the NSA had flagged was now a tab on a soldier's screen,
and with 3 million people able to ask it something,
the number of answers nobody checks gets very large.
An unchecked chatbot answer came close
to putting American troops on a Chinese ship.
During the Iran war, CNN sources say, a U.S. military intelligence
analyst from Special Operations Command in one account
used an AI chatbot to combine open shipping data
with classified intelligence about a Chinese ship in the Middle East.
The chatbot concluded the ship was carrying components
for a nuclear weapons program.
CNN broke the story months later.
This was false.
By the time anyone noticed,
military aircraft were in the air and boarding teams were getting
ready. It was called off
when officials realized the report had been written by AI.
The analyst, CNN sources said, had even let AI do the formatting.
CNN sources said comparable
mistakes had happened before
and that there are no standardized, joint guidelines for
catching AI errors like this. Nobody has said which chatbot it was.
Senator Gillibrand has said decisions over lives must always
be made by human beings,
not unaccountable machines.
And that day, the only thing between a chatbot's report
and a boarding team on a Chinese deck
was a person who noticed, in time.
That near-miss was an accident,
and members of Congress have been shown what it looks
like when someone tricks an AI, on purpose, in a war with China.
In July, Defense One reported on a war game the Washington
think tank CSIS ran with members of two House committees,
and it began with China launching
a massive military exercise around Taiwan in the summer of 2027.
One of its attacks was aimed at an AI system.
A hidden instruction slipped in to make it
do something its users never asked for.
Chinese hackers also went after the ports, freight rail,
and airports America would use to move its forces,
and reporters were asked to leave the room.
Representative McCaul said,
And so cyber is always going to be kind
of the leader in any kinetic war.
And the game also ran an AI misinformation campaign.
U.S. spy agencies had assessed this year, the report noted,
that China isn't planning to invade Taiwan in 2027.
The army in the Pacific is planning for 2027,
anyway, and its commander, General Ronald Clark,
calls the old warning that China could be ready to take Taiwan by then,
still valid. And the man
who owns Grok is helping study what that war gets fought with.
At Quantico this fall, Hegseth launched Project Meridian,
a study of the capabilities
America needs for its next wars,
meant to identify the domains that we must conquer
and capabilities we must master.
Project Meridian will be co-led by three of our nation's
best minds, Elon Musk, Palmer Luckey,
and Newt Gingrich, with findings due early next year.
Hegseth used the same speech to announce
a new autonomous warfare command to be a four-star
command in 2027. Three watchdog groups, Crew,
Public Citizen, and Pogo, have sued the Pentagon
over Meridian, naming Hegseth and the Pentagon's technology chief,
Emil Michael. They say its
meetings are closed, it was never announced
in the Federal Register, and it has no rules on
conflicts of interest for a panel where one
of the three leaders owns a company the Pentagon has
has sworn it relies on.
Whatever that panel finds,
the Pentagon has already put its expectations for that war on file,
in paragraph 8 of Stanley's declaration.
In the event of armed conflict
or other exigent circumstances affecting national security,
Department of War projects,
there is a significant likelihood of an immediate
and substantial escalation in AI inference demands for GrokGov models.
If a war comes,
The Pentagon expects to need far more answers from Grok, and fast.
Therefore, Stanley wrote,
preserving XAI's current data center capacity
is a matter of paramount national security.
In the next paragraph, those data centers sit
on the same level as traditional munitions production,
and replacing XAI, he warned,
would likely mean an unsustainable increase in cost
to the department and thus the American taxpayers.
The war that paragraph 8 is written for has
a year in the army's planning, and it's 2027.
Speaking on October 27th,
General Clark described the time until then as the period when they're
snapping the chalk line to say we need to be ready by this date.
Stanley swore to a likely surge in demand,
and it would run through
a machine whose power plant sits in a live lawsuit.
More of the answers about what gets hit and how fast would come from Grok