ALETH / AI-BRIEF / 2026-03-28 / POWERFUL MYTHOS MODEL REVEALED IN DATA LEAKThe Aleth Briefs link to the sources behind the stories and show how the week unfolded.
The week in five lines:
Anthropic confirmed testing a new model after a leak named it Claude Mythos.
A judge blocked part of the Pentagon’s supply chain risk label on Anthropic.
OpenAI is shutting Sora, and Disney walked away from a planned $1bn stake.
Arm began selling its own chip, the AGI CPU, with Meta as its lead customer.
OpenAI won $10bn more in commitments, taking its record round past $120bn.
Browse by day:
Weekend: Musk Terafab
Monday: OpenAI PE terms · IWF · Epoch maths
Tuesday: Sora · OpenAI $10bn · Arm · TurboQuant · Baltimore · LiteLLM
Wednesday: Manus · GPU smuggling · ARC-AGI-3 · GitHub
Thursday: Mythos · Anthropic ruling · EU AI Act · US Army · Shield AI · Minab
Friday: SoftBank facility
Weekend (21-22 March)
Elon Musk launched Terafab, a joint venture of Tesla, SpaceX and xAI to build a chip fab in Austin, Texas.
Musk presented the plan on Saturday in a livestream on X. He said current chipmaking capacity covers about 2% of what Tesla and SpaceX will need: “We either build the Terafab or we don’t have the chips.” Tesla is targeting a 2-nanometre process at the North Campus of Giga Texas.
TERAFAB Will Build More Compute Than Entire World (Supercut) · Tesla Jigsaw
Tesla and SpaceX announce $25B ‘Terafab’ chip factory · Electrek
Elon Musk announces Terafab project · Engadget
Monday 23 March
OpenAI is offering private equity firms a guaranteed minimum return of 17.5% to join enterprise joint ventures, Reuters reported.
Participating firms would also get early access to new OpenAI models. Anthropic is pursuing a similar venture without a comparable guarantee.
OpenAI sweetens private equity pitch amid turf war with Anthropic · Reuters
The UK’s Internet Watch Foundation (IWF) assessed 8,029 AI-generated images and videos as realistic child sexual abuse material in 2025.
It identified 3,443 AI-generated abuse videos, against 13 in 2024, and classed 65% of the videos as Category A, the most extreme classification.
AI CSAM Report 2026: Harm Without Limits · IWF
Epoch AI said GPT-5.4 Pro (OpenAI) solved an open Ramsey-style problem on hypergraphs from its FrontierMath: Open Problems set.
Kevin Barreto and Liam Price elicited the solution, and problem contributor Will Brian confirmed it. Brian will write it up for publication.
A Ramsey-style Problem on Hypergraphs · Epoch AI
Tuesday 24 March
OpenAI is shutting down its Sora video app, and Disney walked away from a planned $1bn stake.
The Sora team announced on X on Tuesday afternoon that it is “saying goodbye to Sora”. Timelines for the app and the API will follow. Chief financial officer Sarah Friar told CNBC that the company is “facing a lack of compute”. Disney ended its OpenAI partnership. The three-year licence would have let Sora use >200 Disney, Marvel, Pixar and Star Wars characters. Disney said it respects “OpenAI’s decision to exit the video generation business”.
We’re saying goodbye to Sora · Sora on X
OpenAI Shuts Down Sora AI Video, Disney Drops Planned $1B Investment · Variety
OpenAI secured another $10bn in commitments, taking its record round past $120bn.
Friar said Andreessen Horowitz, D.E. Shaw Ventures, MGX, TPG, T. Rowe Price and Microsoft are in the new tranche. The round’s initial target was $100bn. Asked about an IPO, Friar said OpenAI is “starting to build that outcome”.
OpenAI secures an extra $10 billion in record funding round, CFO Friar says · CNBC
Arm began selling its own chip, the AGI CPU, with Meta as lead customer.
Arm says it is delivering its own silicon products for the first time in its >35-year history. It claims >2x the performance per rack of the latest x86 systems. Meta co-developed the chip and will build multiple generations of CPUs with Arm. Other launch partners include OpenAI, Cerebras, Cloudflare and SK Telecom.
Announcing Arm AGI CPU · Arm Newsroom
Meta Partners With Arm to Develop New Class of Data Center Silicon · Meta
Google Research published TurboQuant, which it says compresses the key-value cache of large language models to 3 bits with no loss of accuracy in its tests.
Google reports at least a 6x cut in key-value memory. At 4 bits, attention runs up to 8x faster on Nvidia H100 GPUs than with unquantised 32-bit keys.
TurboQuant: Redefining AI efficiency with extreme compression · Google Research
Baltimore sued X, xAI and SpaceX, alleging that Grok’s sexualised deepfakes breach the city’s consumer protection law.
The complaint, filed in the Circuit Court for Baltimore City, says the companies marketed Grok as a safe assistant without disclosing that it could generate explicit deepfakes. The city seeks civil penalties, an injunction, restitution and disgorgement.
City of Baltimore Sues Over Grok AI’s Role in Generating Deepfakes · DiCello Levitt
Two releases of LiteLLM, an open-source gateway for calling AI model APIs, were published to PyPI with an apparent credential stealer inside.
Versions 1.82.7 and 1.82.8 were live for about 40 minutes on Tuesday before PyPI quarantined them. The maintainers believe the attack came through the Trivy scanner in their build pipeline. Users of the LiteLLM Proxy Docker image were not affected.
Security Update: Suspected Supply Chain Incident · LiteLLM
Wednesday 25 March
China barred Manus co-founders Xiao Hong and Ji Yichao from leaving the country while it reviews Meta’s $2bn purchase, the FT reported.
Beijing is examining whether the sale broke laws on technology exports and outbound investment. Meta had moved >100 Manus staff into its Singapore office in early March. Meta said the deal “complied fully with applicable law”.
China reviews $2bn Manus sale as founders barred from leaving · Financial Times
Beijing’s intervention on Meta’s Manus renews debates on ‘China shedding’ · CNBC
The US Justice Department charged three people with conspiring to smuggle Nvidia AI chips to China through companies in Thailand.
Prosecutors say Tommy Shad English, posing as a Thai buyer, ordered 750 Supermicro servers for about $170m in 2023, 600 of them with export-controlled chips. Stanley Yi Zheng of Hong Kong and Matthew Kelly are also charged. It follows last week’s charges against a Supermicro co-founder.
Three charged with smuggling AI technology to China · US Justice Department
Three more charged with trying to smuggle GPUs to China · The Register
The ARC Prize Foundation launched ARC-AGI-3, an interactive benchmark of game-like environments, with $2m in prizes.
Humans with no instructions solved every environment. On the leaderboard, which tests models via API without custom scaffolding, every frontier model scored <1%.
GitHub will train its AI models on Copilot interaction data from Free, Pro and Pro+ users from 24 April unless they opt out.
The data covers inputs, outputs, code snippets and associated context. Copilot Business and Copilot Enterprise users are not affected.
Updates to GitHub Copilot interaction data usage policy · GitHub
Thursday 26 March
Anthropic confirmed it is testing a new model, which a leaked draft calls Claude Mythos.
A draft blog post, left in a publicly searchable data store, named the model and called it “by far the most powerful AI model we’ve ever developed”. Anthropic told Fortune the model is “a step change” and is being trialled by early access customers. The draft describes a new tier of model larger than Anthropic’s Opus line. It says the model is “currently far ahead of any other AI model in cyber capabilities”. Anthropic blamed “human error” in configuring its content management system.
Anthropic acknowledges new AI model with ‘step change’ in capabilities · Fortune
Anthropic left details of unreleased model unsecured in security lapse · Fortune
A federal judge blocked part of the Pentagon’s move to label Anthropic a supply chain risk.
Judge Rita Lin wrote that nothing in the statute supports “the Orwellian notion” that an American company may be branded a potential adversary for disagreeing with the government. She called the government’s conduct “classic illegal First Amendment retaliation” and delayed her order by a week to allow an appeal. At Tuesday’s hearing Lin said the Pentagon’s moves look “like an attempt to cripple Anthropic”. Anthropic must still persuade the DC Circuit court.
Order granting injunction · CourtListener · Hearing transcript · CourtListener
‘Premature’: Anthropic still in trouble despite court win · Politico
The European Parliament adopted its position to delay the EU AI Act’s high-risk rules and ban AI nudifier apps.
The vote was 569 to 45, with 23 abstentions. MEPs propose 2 December 2027 for high-risk systems listed in the Act and 2 August 2028 for those covered by EU sectoral law. Negotiations with the Council can now begin.
AI Act: delayed application, ban on nudifier apps · European Parliament
The US Army conditionally selected Carlyle and CyrusOne, owned by KKR and BlackRock funds, to negotiate building commercial data centres on two bases.
Carlyle would build at Fort Bliss, Texas, and CyrusOne at Dugway Proving Ground, Utah, under an Army programme that leases land on its bases. The sites are due to reach initial operating capability in FY2027 and FY2029. The deal is not yet final.
Army reaches conditional agreement for hyperscaled data centers · US Army
Defence technology company Shield AI is raising $2bn at a $12.7bn valuation and plans to buy simulation software maker Aechelon Technology.
The package is a $1.5bn Series G led by Advent International plus $500m of fixed-return preferred equity from Blackstone funds. Aechelon’s software supports the Pentagon’s Joint Simulation Environment.
Shield AI to acquire Aechelon and raise $2B at $12.7B valuation · Shield AI
A Guardian long read argued that the Maven targeting system, not Anthropic’s Claude, lay behind the US strike on a school in Minab, Iran.
The strike on the Shajareh Tayyebeh school on 28 February killed 175-180 people, mostly girls. Kevin T Baker writes that coverage centred on whether Claude chose the target, while “the targeting for Operation Epic Fury ran on a system called Maven”.
AI got blame for Iran school bombing. The truth is far more worrying · The Guardian
Friday 27 March
SoftBank signed a $40bn bridge facility, mainly to fund its $30bn follow-on investment in OpenAI.
The lenders are JPMorgan Chase, Goldman Sachs, Mizuho Bank, Sumitomo Mitsui and MUFG Bank. The facility carries no collateral and matures on 25 March 2027.
Bridge Facility Agreement for the Follow-on Investments in OpenAI · SoftBank
ALETH · OPEN ACCESS RESEARCH