DoD freezes out Anthropic, Musk’s emails sink his case, Meta drops 9.4%

TL;DR

Pentagon signed classified-network AI deals with seven vendors and excluded Anthropic, branding it a ‘supply chain risk to national security’ over a refused contract clause.
Week 1 of Musk v. Altman damaged Musk more than OpenAI: his own 2018 emails surfaced and Microsoft used opening day to publicly amend its OpenAI partnership.
Meta acquired humanoid startup ARI and bumped 2026 capex to $125–145B; shares fell ~9.4% on the week despite still having no robot to ship.
Minnesota became the first US state to ban AI nudification apps, with developer fines up to $500K, as fresh CSAM evidence implicates xAI’s Grok.
Microsoft put a Legal Agent inside Word for contract review, betting structured legal workflows will earn lawyer trust where general LLM prompting hasn’t.

Today’s AI news lives in the verdicts coming back from outside the labs. The Pentagon signed classified-network deals with seven AI vendors and pointedly excluded Anthropic, branding the company a “supply chain risk to national security” — a designation usually reserved for firms like Huawei — after Anthropic refused an “any lawful use” clause the others accepted. A federal courtroom spent week one of Musk v. Altman dismantling Musk’s case with his own 2018 emails and an xAI distillation admission, while Microsoft and OpenAI used opening day to visibly loosen the partnership Musk is suing over. And Wall Street took 9.4% off Meta’s market cap after a $10B capex hike landed alongside a humanoid-AI acquisition that left Meta still without a robot to ship.

Three different institutions — military procurement, federal court, public markets — each handed an AI company a verdict that contradicted its own framing, and in two of the three the company supplied the evidence itself. The roundup covers Minnesota’s first-in-the-nation nudification ban, Microsoft’s Legal Agent for Word, and more.

Pentagon brands Anthropic a supply-chain risk over safety terms

Source: the-verge-ai · published 2026-05-01

TL;DR

The Pentagon signed classified-network AI deals with OpenAI, Google, Microsoft, AWS, Nvidia, xAI and Reflection — Anthropic was excluded after refusing an “any lawful use” clause.
OpenAI and Microsoft signed the same headline terms only after securing non-binding side letters on human oversight; Anthropic called that safety theater.
DoD labeled Anthropic a “supply chain risk to national security,” a designation previously reserved for firms like Huawei.
Reflection AI — a $20–25B startup with no public model and ties to 1789 Capital — got peer-tier classified access alongside the hyperscalers.

The fault line was a contract clause, not a snub

The Verge and TechCrunch framed Friday’s announcement as Anthropic being “left out.” The fuller record shows a deliberate refusal. Defense Secretary Pete Hegseth’s procurement template requires vendors to permit “any lawful use” of their models on classified networks. Anthropic insisted on explicit contractual prohibitions against mass domestic surveillance and fully autonomous lethal weapons. Pentagon CTO Emil Michael called those carve-outs a “private company veto” over national-security decisions and walked ¹.

The Pentagon designated Anthropic a “supply chain risk to national security” — the first time such a label, historically reserved for foreign adversaries like Huawei, was applied to an American firm ².

That escalation matters more than the contract loss. Compliance analysts flagged the label as a category jump — a tool built for foreign adversaries now aimed at a domestic AI lab ². CFO filings reportedly warn the freeze-out could cost Anthropic “multiple billions” in 2026 revenue ³.

The side-letter compromise everyone else took

The most under-reported detail: OpenAI and Microsoft signed the same “lawful use” language Anthropic rejected, then negotiated separate, non-binding assurances on human oversight after the fact ⁴. Anthropic refused that structure on the grounds that a side letter without contractual force is cosmetic. The cluster therefore isn’t “Anthropic vs. seven vendors” — it’s “Anthropic refused a compromise the others quietly accepted.”

flowchart TB
    A["Hegseth 'any lawful use' clause"] --> B{Vendor response}
    B -->|"Sign + non-binding<br/>side letter"| C["OpenAI, Microsoft,<br/>Google, AWS, Nvidia, xAI"]
    B -->|"Sign without<br/>side letter"| D["Reflection AI<br/>(no public model)"]
    B -->|"Refuse: demand<br/>contractual carve-outs"| E[Anthropic]
    C --> F[Classified-network access]
    D --> F
    E --> G["'Supply chain risk'<br/>designation"]

Reflection’s inclusion is the procurement story

Reflection AI sits on the winners list with no publicly released model, a $20–25B valuation, and backing from 1789 Capital — the venture firm where Donald Trump Jr. is a partner ⁵. Its founders bring DeepMind and AlphaGo pedigree, but the company has shipped no benchmarks or weights that would justify peer-tier classified access alongside hyperscalers running production models for hundreds of millions of users. The Verge piece doesn’t mention this; the procurement-integrity question it raises is arguably bigger than the Anthropic exclusion.

What’s actually at stake

Two threads suggest the consensus is shakier than the announcement reads. Dario Amodei met White House officials in early May about “Mythos,” Anthropic’s cybersecurity model, which Michael himself called a “separate national security moment” — meaning the ban is porous where the capability is unique ³. And the deal landed against visible labor dissent: 600+ Google employees signed a protest letter, and a “QuitGPT” movement claims roughly 1.5 million accounts ⁶. Vendor press releases mention neither.

The real precedent here isn’t the contract values. It’s the Pentagon establishing that enforceable safety guardrails — as opposed to revocable assurances — are now a procurement disqualifier, and that six of seven major AI vendors are willing to operate under that rule.

Musk’s own emails and xAI admission sink his OpenAI case

Source: mit-tech-review-ai · published 2026-05-01

TL;DR

Week 1 of Musk v. Altman produced more damage to Musk than to OpenAI: an xAI distillation admission, inflated funding claims, and his own 2018 emails.
Judge Gonzalez Rogers barred “doomsday” AI-safety arguments and made the jury advisory only — she decides liability and remedy herself.
Legal scholars expect at best ~$44M restitution, not the $150B structural unwinding Musk seeks.
Microsoft and OpenAI used opening day to amend their partnership, visibly loosening the alliance Musk is suing over.

The distillation admission isn’t a one-off

The line MIT Tech Review buried — Musk conceding on the stand that xAI distills OpenAI’s models as “standard industry procedure” — is the most consequential moment of the week, because it isn’t isolated. Anthropic revoked xAI’s Claude API access in January 2026 after catching xAI engineers routing Claude through the Cursor IDE to accelerate Grok development ⁷. Two frontier labs, same pattern, same company. Musk’s framing is simultaneously true (every lab polices distillation) and self-incriminating (every lab’s ToS forbids it). It hands OpenAI its cleanest narrative: this is a competitor’s lawsuit dressed as a fiduciary one.

The judge is quietly gutting the theatrics

A four-part cluster of trial coverage this week converged on the same observation: Judge Yvonne Gonzalez Rogers is running a much narrower trial than Musk’s testimony suggests. The nine-person jury’s verdict is advisory only — she retains final authority on liability and remedies — and she has instructed counsel to avoid “doomsday talk,” ruling that the case is about contract law and fiduciary duties, not AI extinction risk ⁸. That single procedural ruling neutralizes roughly half of Musk’s rhetorical strategy. Columbia’s Eric Talley argues that using “relatively informal” text messages to establish a tens-of-millions-of-dollars contract “stacks the deck” against Musk, and that even a liability win likely produces disgorgement of his ~$44M contribution rather than any structural overhaul of OpenAI ⁹.

The receipts cut against Musk

OpenAI walked into trial with documentary counter-evidence prepared. The most-cited exhibit is a 2018 email exchange in which Musk agrees with then–chief scientist Ilya Sutskever:

it’s totally OK to not share the science

— directly undermining his courtroom claim that openness was a binding founding commitment ¹⁰. Other released emails show Musk himself attempting to take majority control and the CEO role during the same restructuring he now characterizes as theft ¹⁰. The trial record also pinned his actual cash contribution at roughly $38 million, well below the $1 billion figure he has repeated in public for years ¹¹.

Musk’s public narrative	Week 1 evidence
Funded OpenAI with ~$1B	~$38M actual contribution ¹¹
Defended openness as founding principle	2018 email: “totally OK to not share the science” ¹⁰
Opposed for-profit capture	Sought majority control + CEO role himself ¹⁰
xAI competes on the merits	Admits distilling OpenAI; cut off by Anthropic for same ⁷

The Microsoft pivot, timed to opening day

On the trial’s opening day in April 2026, Microsoft and OpenAI announced a significant amendment to their partnership: Microsoft’s IP license becomes non-exclusive, and OpenAI is now free to use AWS and Google Cloud ¹². The timing was not coincidental. Musk’s complaint leans heavily on the theory that Microsoft captured a nonprofit; the amendment lets OpenAI’s lawyers point at a visibly loosened alliance whenever that argument surfaces.

What’s actually at stake

Strip away the warnings about extinction and the “I was duped” testimony, and week 1 produced a coherent picture: a trial where the bench has narrowed the questions to contract and fiduciary duty, where the documentary record cuts against the plaintiff, and where the most probable outcome — even on a Musk win — is restitution of his ~$44M contribution rather than the $150B structural unwinding he seeks ⁹. The headline risk to OpenAI this week was reputational. The legal risk shrank.

Meta buys humanoid startup ARI; capex hike sinks stock 9.4%

Source: techcrunch-ai · published 2026-05-01

TL;DR

Meta acquired Assured Robot Intelligence to license a cross-vendor humanoid intelligence stack, not to build its own robot.
The deal landed alongside a $10B capex bump to $125–145B for 2026; Meta shares fell ~9.4% on the week.
ARI’s real assets are two embodied-AI researchers and e-Flesh, a 3D-printable magnetic tactile skin.
Figure and Tesla already have humanoids on factory floors; Meta has none, which undercuts its “Android of humanoids” pitch.

The pitch: be the OS, not the OEM

Meta’s acquisition of Assured Robot Intelligence (ARI) is being framed by people close to the deal as a play to become the “Android of humanoids” — the intelligence-and-sensor layer third-party hardware makers license rather than a vertically integrated competitor to Tesla Optimus or Figure ¹³. ARI’s founders, Lerrel Pinto (NYU, ex-Fauna Robotics, which Amazon bought earlier in 2026) and Xiaolong Wang (UCSD, ex-Nvidia), are being folded into Meta Superintelligence Labs alongside the “Metabot” reference design run by ex-Cruise CEO Marc Whitten ¹³.

The technical asset that justifies the price tag is e-Flesh (and its sibling AnySkin): a low-cost, 3D-printable magnetic tactile sensor designed to generalize across robot instances without per-unit recalibration ¹⁴. That plugs directly into Meta’s existing Sparsh/Digit 360 tactile research and gives it a credible answer to contact-rich manipulation — the part of robotics where vision-language models are weakest.

The market reaction was ugly

Wall Street did not reward the strategy. Meta paired the ARI announcement with a $10 billion increase to its 2026 capex guidance, now $125–145B, and the stock dropped roughly 9.4% over the announcement week ¹⁵. The deal also closed two weeks after Meta laid off about 8,000 employees — a juxtaposition LiveMint put directly in its headline ¹⁶ and one Zuckerberg has defended as the price of funding physical-AI ambitions.

There’s a regulatory subplot too. Senators Warren and Blumenthal have asked the FTC and DOJ to scrutinize “reverse acquihires” — the structure Meta used for its $14.3B Scale AI stake — as de facto mergers engineered to dodge HSR review ¹⁷. ARI is small enough that this specific deal will likely escape, but it fits the pattern regulators are now actively watching.

The “Android” analogy has a deployment problem

The OS-for-humanoids framing obscures how far behind Meta is on the only metric that ultimately matters: robots doing real work. Figure’s Helix 02 is running pilots at BMW’s Spartanburg plant, where it recently demonstrated a 4-minute autonomous dishwasher-unloading cycle ¹⁸. Tesla Optimus is deployed inside Tesla’s own factories. Meta has no comparable fleet.

The Android analogy depends on a data moat from millions of deployed units, not from social-graph scale.

That’s the load-bearing critique ¹⁴. Android won because OEMs shipped phones that generated training and telemetry data Google could compound on. Meta is trying to sell an intelligence layer before any humanoid platform — its own or a partner’s — is shipping at volume.

Net read

ARI gives Meta real tactile IP and two of the stronger embodied-AI researchers on the market. But the deal is being read by analysts as a defensive talent grab inside a capex story investors are already punishing ¹⁵¹⁶, executed through a structure regulators are starting to dislike ¹⁷, aimed at a market where the vertically integrated incumbents are years ahead on the part that counts ¹⁸. The “Android of humanoids” line is a good slogan. It is not yet a product.

Round-ups

Minnesota passes ban on fake AI nudes; app makers risk $500K fines

Source: ars-technica-ai

Minnesota becomes the first US state to ban nudification apps, with developers facing fines up to $500,000. The law arrives alongside fresh evidence of CSAM generated via xAI’s Grok, intensifying pressure on platforms hosting image-generation tools.

Microsoft wants lawyers to trust its new AI agent in Word documents

Source: the-verge-ai

Microsoft’s new Legal Agent inside Word targets contract review and negotiation history for legal teams, following structured workflows modeled on legal practice rather than relying on general-purpose LLM prompting to interpret commands.

Replit’s Amjad Masad on the Cursor deal, fighting Apple, and why he’d rather not sell

Source: techcrunch-ai

Replit CEO Amjad Masad fielded acquisition questions at TechCrunch’s StrictlyVC event after reports that rival Cursor is in talks to be bought by SpaceX for $60 billion. Masad signaled he’d rather keep Replit independent and discussed the company’s posture toward Apple and Anthropic.

Meta cuts contractors who reported seeing Ray-Ban Meta users have sex

Source: ars-technica-ai

Meta terminated a Kenya-based contractor team that reviewed Ray-Ban Meta smart glasses footage after workers reported seeing wearers having sex and other intimate moments. Meta said the contractors did not meet its standards, sidestepping the underlying privacy questions about always-on camera review.

Source: mit-tech-review-ai

A T-Mobile-based MVNO marketed to Christians launches next week with network-level blocking of pornography that account holders cannot disable, plus filters for gender-related content. Security experts say it’s the first US carrier to enforce such blocks at the network layer.

Christian content creators are outsourcing AI slop to gig workers on Fiverr

Source: the-verge-ai

Fiverr freelancers are advertising rapid-turnaround AI-generated Bible videos for Christian YouTube and TikTok creators, turning the gig platform once known for specialized creative labor into a pipeline for outsourced generative-AI religious content.

MIT Tech Review EmTech AI conference session writeups

Source: mit-tech-review-ai

MIT Technology Review published session writeups from its EmTech AI conference covering cybersecurity rebuilt around AI-era attack surfaces and the operational challenges of scaling AI deployments while preserving data sovereignty, framing both as failures of layering AI onto legacy stacks.

Further reading:

Operationalizing AI for Scale and Sovereignty — mit-tech-review-ai

Military Times — https://www.militarytimes.com/news/pentagon-congress/2026/05/01/pentagon-freezes-out-anthropic-as-it-signs-deals-with-ai-rivals/

Anthropic refused the Pentagon’s ‘all lawful use’ contract clause, insisting on contractual prohibitions against using Claude for mass domestic surveillance and fully autonomous lethal weapons — terms Pentagon CTO Emil Michael characterized as an unacceptable ‘private company veto’ over national security decisions.

↩
Ethixbase360 compliance analysis — https://ethixbase360.com/compliance-implications-of-anthropics-dispute-with-the-pentagon/

The Pentagon designated Anthropic a ‘supply chain risk to national security’ — the first time such a label, historically reserved for foreign adversaries like Huawei, was applied to an American firm.

↩ ↩²
TipRanks — ‘Pentagon Shifts Focus to Mythos AI’ — https://www.tipranks.com/news/pentagon-shifts-focus-to-mythos-ai-says-anthropic-still-blacklisted

Dario Amodei met with White House officials in early May to discuss ‘Mythos,’ Anthropic’s cybersecurity model; Pentagon CTO Emil Michael called it a ‘separate national security moment’ that could bypass the broader ban. Internal CFO filings warned exclusion could cost Anthropic ‘multiple billions of dollars’ in 2026 revenue.

↩ ↩²
Gizmodo — ‘The Rest of Big Tech Piles In to Take the Pentagon Deal That Anthropic Wouldn’t’ — https://gizmodo.com/the-rest-of-big-tech-piles-in-to-take-the-pentagon-deal-that-anthropic-wouldnt-2000753280

OpenAI and Microsoft signed the broader ‘lawful use’ terms only after securing separate, non-binding side assurances regarding human oversight — a compromise Anthropic explicitly rejected as safety theater.

↩
SiliconAngle — https://siliconangle.com/2026/05/01/pentagon-inks-ai-procurement-deals-seven-companies-leaves-anthropic/

Reflection AI — a startup with no publicly released model — was included alongside OpenAI, Google and Nvidia despite shipping no public product, carrying a $20–25 billion valuation and backing from 1789 Capital, the venture firm where Donald Trump Jr. is a partner.

↩
Times of India — citing Google employee letter — https://timesofindia.indiatimes.com/technology/tech-news/pentagon-signs-classified-ai-deal-that-google-employees-objected-to-with-seven-tech-companies-heres-the-full-list-as-war-department-says-these-strategic-partners-share/articleshow/130708959.cms

Over 600 Google employees signed a protest letter objecting to the classified-network deal, and a ‘QuitGPT’ user movement of roughly 1.5 million accounts emerged in response — signaling internal labor dissent that vendor press releases omit.

↩
The Rundown AI — https://www.therundown.ai/p/anthropic-pulls-plug-on-xais-claude-access

Anthropic revoked xAI’s access to the Claude API after discovering that xAI engineers were using the models — specifically through the Cursor IDE — to accelerate their internal development

↩ ↩²
GeekWire — inside the courthouse — https://www.geekwire.com/2026/musk-v-altman-inside-the-courthouse-as-microsofts-13-billion-openai-bet-goes-on-trial/

the nine-person jury’s verdict will be advisory only; she retains the final authority to decide liability and any potential remedies… the judge instructed attorneys to avoid ‘doomsday talk,’ ruling that the trial is about contract law and fiduciary duties, not the existential safety risks of artificial intelligence

↩
Duke Law / Daily Journal analysis (Prof. Eric Talley) — https://law.duke.edu/web/sites/default/files/pdf/Daily_Journal-Open_AI.pdf

using informal texts to establish a tens-of-millions-of-dollars contract is ‘relatively informal’ and ‘stacks the deck’ against Musk… a more likely outcome, should Musk win on liability, is ‘disgorgement’ or restitution of his $44 million investment rather than a structural overhaul

↩ ↩²
SF Chronicle — OpenAI’s released emails — https://www.sfchronicle.com/tech/article/openai-fires-back-musk-suit-publishes-emails-18707673.php

a 2018 exchange shows Musk agreeing with Chief Scientist Ilya Sutskever that ‘it’s totally OK to not share the science’ as the technology advances… Musk attempted to gain majority control and CEO status during the restructuring himself

↩ ↩² ↩³ ↩⁴
WKZO / AP wire — Musk testimony takeaways — https://wkzo.com/2026/05/01/key-takeaways-from-musks-testimony-at-openai-trial/

Musk’s total financial contribution was roughly $38 million — significantly lower than the $1 billion figure Musk had previously cited in public narratives

↩ ↩²
GeekWire — Microsoft/OpenAI partnership amendment — https://www.geekwire.com/2026/microsoft-and-openai-revamp-partnership-with-trial-in-elon-musk-suit-set-to-begin/

On the opening day of the trial in April 2026, Microsoft and OpenAI announced a significant amendment to their partnership, making Microsoft’s intellectual property license non-exclusive and allowing OpenAI to use other cloud providers like AWS and Google Cloud

↩
LetsDataScience — https://letsdatascience.com/news/meta-acquires-assured-robot-intelligence-for-humanoid-ai-32e71601

Meta’s interest in ARI reflects a strategy to become the ‘Android of humanoids,’ providing the underlying intelligence layer for various third-party hardware manufacturers

↩ ↩²
AwesomeAgents (HN/X roundup) — https://awesomeagents.ai/news/meta-ari-humanoid-physical-agi/

ARI’s flagship technology, e-Flesh, is a low-cost, 3D-printable magnetic tactile sensor that allows robots to handle delicate objects like eggs and plushies with human-like dexterity

↩ ↩²
MLQ.ai analyst note — https://mlq.ai/news/meta-acquires-robotics-startup-assured-robot-intelligence-ari-for-humanoid-robot-development/

Meta’s stock fell 9.4% over the week of the announcement… the acquisition was paired with a $10 billion hike in 2026 capital expenditure guidance, now projected at $125–$145 billion

↩ ↩²
LiveMint — https://www.livemint.com/companies/news/meta-acquires-humanoid-robot-startup-assured-robot-intelligence-weeks-after-announcing-8-000-layoffs/amp-11777692898608.html

Meta acquires humanoid robot startup Assured Robot Intelligence weeks after announcing 8,000 layoffs

↩ ↩²
Credit and Collection News (on Sen. Warren letter) — https://www.creditandcollectionnews.com/sen-warren-others-urge-ftc-doj-to-scrutinize-tech-ai-acquihire-deals-for-antitrust-violations/

Senators Warren and Blumenthal urged the FTC and DOJ to investigate ‘reverse acquihires,’ characterizing them as de facto mergers designed to evade standard regulatory reviews

↩ ↩²
humanoid.guide leaderboard — https://humanoid.guide/figure-helix/

Figure currently holds a significant lead in commercial validation; its robots are already conducting pilots at BMW’s Spartanburg plant, where Helix 02 recently demonstrated a 4-minute autonomous dishwasher unloading cycle

↩ ↩²

DoD freezes out Anthropic, Musk's emails sink his case, Meta drops 9.4%

DoD freezes out Anthropic, Musk’s emails sink his case, Meta drops 9.4%

TL;DR

Pentagon brands Anthropic a supply-chain risk over safety terms

TL;DR

The fault line was a contract clause, not a snub

The side-letter compromise everyone else took

Reflection’s inclusion is the procurement story

What’s actually at stake

Further reading

Musk’s own emails and xAI admission sink his OpenAI case

TL;DR

The distillation admission isn’t a one-off

The judge is quietly gutting the theatrics

The receipts cut against Musk

The Microsoft pivot, timed to opening day

What’s actually at stake

Further reading

Meta buys humanoid startup ARI; capex hike sinks stock 9.4%

TL;DR

The pitch: be the OS, not the OEM

The market reaction was ugly

The “Android” analogy has a deployment problem

Net read

Round-ups

Minnesota passes ban on fake AI nudes; app makers risk $500K fines

Microsoft wants lawyers to trust its new AI agent in Word documents

Replit’s Amjad Masad on the Cursor deal, fighting Apple, and why he’d rather not sell

Meta cuts contractors who reported seeing Ray-Ban Meta users have sex

Christian content creators are outsourcing AI slop to gig workers on Fiverr

MIT Tech Review EmTech AI conference session writeups

Jack Sun, writing.

DoD freezes out Anthropic, Musk’s emails sink his case, Meta drops 9.4%

TL;DR

Pentagon brands Anthropic a supply-chain risk over safety terms

TL;DR

The fault line was a contract clause, not a snub

The side-letter compromise everyone else took

Reflection’s inclusion is the procurement story

What’s actually at stake

Further reading

Musk’s own emails and xAI admission sink his OpenAI case

TL;DR

The distillation admission isn’t a one-off

The judge is quietly gutting the theatrics

The receipts cut against Musk

The Microsoft pivot, timed to opening day

What’s actually at stake

Further reading

Meta buys humanoid startup ARI; capex hike sinks stock 9.4%

TL;DR

The pitch: be the OS, not the OEM

The market reaction was ugly

The “Android” analogy has a deployment problem

Net read

Round-ups

Minnesota passes ban on fake AI nudes; app makers risk $500K fines

Microsoft wants lawyers to trust its new AI agent in Word documents

Replit’s Amjad Masad on the Cursor deal, fighting Apple, and why he’d rather not sell

Meta cuts contractors who reported seeing Ray-Ban Meta users have sex

A new US phone network for Christians aims to block porn and gender-related content

Christian content creators are outsourcing AI slop to gig workers on Fiverr

MIT Tech Review EmTech AI conference session writeups

Footnotes

Jack Sun, writing.