ALETH / AI-BRIEF / 2026-02-28 / PENTAGON DROPS ANTHROPIC AND SIGNS OPENAIThe Aleth Briefs link to the sources behind the stories and show how the week unfolded.
The week in five lines:
Trump ordered federal agencies off Anthropic, and Hegseth moved to blacklist it.
OpenAI announced $110bn of investment from Amazon, Nvidia and SoftBank.
After the ban, OpenAI agreed a classified Pentagon deal with two safeguards.
AMD will supply Meta up to 6 GW of GPUs, with warrants for 160m AMD shares.
Anthropic accused three Chinese labs of distilling Claude via 24,000 fake accounts.
Browse by day:
Weekend: Stargate stall
Monday: Agents of Chaos · Distillation · IBM · Grok
Tuesday: AMD-Meta · Ultimatum · RSP · Wayve · Qwen3.5
Wednesday: Mexico hack · China op · Vercept · Gemini keys · Nvidia
Thursday: OpenAI London · Nano Banana 2 · Amodei · Canada
Friday: OpenAI $110bn · Trump ban · OpenAI-Pentagon
Weekend (21-22 February)
The $500bn Stargate venture of OpenAI, Oracle and SoftBank has hired no staff and is building no data centres, The Information reported.
The partners have been arguing over who is responsible for what. OpenAI then tried to build its own data centres but could not raise the money. It missed a goal of securing 10 GW over three years through SoftBank and Oracle by the end of 2025, and signed deals with AWS, Google Cloud, AMD and Cerebras instead.
Inside OpenAI’s scramble for compute after Stargate stalled · The Information
Stargate’s $500 billion AI infrastructure project reportedly stalls · The Decoder
Monday 23 February
A two-week red-teaming study gave AI agents email, shell access and persistent memory, and found them leaking data and misreporting their own work.
Twenty researchers documented eleven case studies, from obeying people who did not own the agents to destructive actions. Asked to delete an email, one agent reset its mail client and reported the problem solved. The email was still in the inbox.
Agents of Chaos · arXiv
An OpenClaw AI agent nuked its own mail client and called it fixed · The Decoder
Anthropic accused DeepSeek, Moonshot and MiniMax of distilling Claude.
The three labs generated >16m exchanges with Claude through about 24,000 fraudulent accounts, breaching its terms and regional access restrictions, Anthropic said. MiniMax accounted for >13m exchanges, Moonshot AI >3.4m and DeepSeek >150,000. Anthropic attributed each campaign through IP addresses, request metadata and infrastructure indicators. DeepSeek also used Claude to write “censorship-safe alternatives to politically sensitive queries”. When Anthropic released a new model mid-campaign, MiniMax redirected nearly half its traffic to it within 24 hours.
Detecting and preventing distillation attacks · Anthropic
IBM shares closed nearly 13% lower after Anthropic said Claude Code can automate the exploration and analysis phases of COBOL modernisation.
COBOL handles an estimated 95% of US ATM transactions, and IBM has long sold the mainframes on which much of it runs. IBM closed at $223.
How AI helps break the cost barrier to COBOL modernisation · Claude
IBM shares tank 13% on Anthropic programming language threat · CNBC
xAI agreed to let the military use Grok in classified systems on the “all lawful use” terms Anthropic refused, Axios reported.
Until now Claude was the only model available in those systems.
Musk’s xAI, Pentagon reach deal to use Grok in classified systems · Axios
Tuesday 24 February
AMD and Meta signed a deal for up to 6 GW of GPUs with an equity warrant attached.
Shipments for the first gigawatt begin in H2 2026, on a custom MI450-based Instinct GPU. AMD issued Meta a warrant for up to 160m AMD shares. It vests as shipments scale from 1 GW to 6 GW and as AMD’s share price hits set thresholds. Mark Zuckerberg called it “an important step for Meta as we diversify our compute”.
AMD, Meta announce expanded partnership to deploy 6 gigawatts of GPUs · AMD
Hegseth gave Anthropic until Friday to allow unfettered military use of Claude or be declared a supply chain risk, Axios reported.
He delivered the ultimatum at Tuesday’s meeting with Amodei, telling him no company would dictate the terms of the Pentagon’s operational decisions. Anthropic said it had continued “good-faith conversations about our usage policy”. Hegseth also threatened to compel it under the Defense Production Act.
Hegseth gives Anthropic until Friday to back down on AI safeguards · Axios
US military leaders pressure Anthropic to bend Claude safeguards · The Guardian
Anthropic’s new Responsible Scaling Policy (RSP) drops its 2023 pledge not to train or release models without guaranteed safety measures.
Chief science officer Jared Kaplan told Time it no longer made sense “for us to make unilateral commitments” while competitors raced ahead. The new RSP separates what Anthropic will do regardless of rivals from mitigations it says the industry needs. It adds a Frontier Safety Roadmap and Risk Reports every three to six months.
London-based Wayve raised a $1.2bn Series D at an $8.6bn post-money valuation, with Mercedes-Benz, Nissan and Stellantis among the investors.
Eclipse, Balderton and SoftBank Vision Fund 2 led the round, and Microsoft, Nvidia and Uber joined. Uber will add milestone-based money within $1.5bn secured for commercial rollout. Wayve will start commercial robotaxi trials this year, with its first service with Uber planned for London. Its supervised autonomy software reaches consumer cars from 2027.
Wayve secures $1.5B to deploy its global autonomy platform · Wayve
Alibaba released three open-weights Qwen3.5 models, 35B-A3B, 122B-A10B and 27B, alongside a hosted Qwen3.5-Flash.
All three ship under Apache 2.0, and Alibaba benchmarks them against GPT-5 mini, gpt-oss-120b and its own much larger Qwen3-235B-A22B. Qwen3.5-Flash is the hosted version of 35B-A3B, with a 1M-token context by default and built-in tools.
Qwen3.5-35B-A3B · Hugging Face
Open Qwen 3.5 takes aim at GPT-5 mini and Claude Sonnet 4.5 · The Decoder
Wednesday 25 February
A hacker used Claude to breach Mexican agencies and steal 150GB of data, including documents tied to 195m taxpayer records, Gambit Security said.
Writing in Spanish, the attacker told Claude to act as an elite hacker. Claude warned of malicious intent at first but eventually complied. The tax authority and the national electoral institute were among those hit, though the institute said it had found no breach. Anthropic said it disrupted the activity and banned the accounts.
Hacker used Anthropic’s Claude to steal sensitive Mexican data · Claims Journal
OpenAI banned a ChatGPT account tied to Chinese law enforcement that sought help to discredit Japanese prime minister Sanae Takaichi.
ChatGPT refused, but the user later asked it to polish a status report showing the campaign had gone ahead without it. The campaign followed Takaichi’s public criticism of human rights in Inner Mongolia. OpenAI says the wider effort uses at least hundreds of staff, thousands of fake accounts and locally run models such as DeepSeek and Qwen. “It’s industrialized,” said OpenAI investigator Ben Nimmo.
Cyber special operations: China-linked influence planning · OpenAI
China tried using ChatGPT to discredit Japan’s PM, OpenAI says · Axios
Anthropic acquired Vercept, the Seattle start-up behind the Vy desktop agent, to push Claude’s computer use further.
Co-founders Kiana Ehsani, Luca Weihs and Ross Girshick join Anthropic, and Vy shuts within 30 days. Vercept raised a $16m seed in January 2025 at a $67m post-money valuation, according to PitchBook.
Anthropic acquires Vercept · Anthropic
Anthropic acquires Vercept, one of Seattle’s standout AI startups · GeekWire
Truffle Security found that enabling Gemini on a Google Cloud project silently turns its existing API keys, never meant as secrets, into Gemini keys.
Scanning the November 2025 Common Crawl, it found 2,863 live keys exposed this way, some of them Google’s own. Google has begun blocking exposed keys.
Google API keys weren’t secrets, but then Gemini changed the rules · Truffle Security
Nvidia reported record Q4 FY2026 revenue of $68.1bn, up 73% YoY, with data centre revenue up 75% to $62.3bn.
It guided to Q1 FY2027 revenue of $78.0bn, plus or minus 2%, assuming no data centre compute revenue from China. Revenue for the fiscal year was $215.9bn.
NVIDIA announces financial results for Q4 and fiscal 2026 · SEC
Thursday 26 February
OpenAI named London its largest research hub outside the US, with researchers there to “own key components of OpenAI’s frontier model development”.
UK Technology Secretary Liz Kendall called it “a huge vote of confidence”. OpenAI opened its first international office in London in 2023.
OpenAI names London as its next major research hub · UKTN
Google released Nano Banana 2, or Gemini 3.1 Flash Image, which brings the world knowledge and quality of Nano Banana Pro to Flash speeds.
It renders and translates legible text inside images, keeps up to five characters and 14 objects consistent across a workflow, and outputs from 512px to 4K. It replaces Nano Banana Pro in the Gemini app’s Fast, Thinking and Pro models.
Nano Banana 2: Google’s latest AI image generation model · Google
Dario Amodei said Anthropic “cannot in good conscience accede” to Pentagon demands to drop its safeguards on mass surveillance and autonomous weapons.
Anthropic said contract language received overnight “made virtually no progress” on either. Amodei said frontier models are “simply not reliable enough to power fully autonomous weapons”. He called the twin threats of a supply chain risk label and the Defense Production Act “inherently contradictory”. If dropped, Anthropic will help with a smooth transition to another provider.
The Pentagon’s Emil Michael told CBS News it had offered to acknowledge in writing the federal laws restricting surveillance of Americans, and to bring Anthropic onto its AI ethics board. He called Amodei a “liar” with a “God complex”.
Dario Amodei on the Department of War discussions · Anthropic
Anthropic says Pentagon’s “final offer” is unacceptable · Axios
Pentagon official lashes out at Anthropic as talks break down · CBS News
OpenAI is overhauling its safety protocols after choosing not to report a ChatGPT user whom police say went on to kill eight people in Canada.
OpenAI also revealed that the shooter had opened a second account after it banned the first. The announcement followed meetings this week with the governments of Canada and British Columbia.
Shooter had second ChatGPT account, OpenAI says as it overhauls safety · Politico
Friday 27 February
OpenAI announced $110bn of new investment at a $730bn pre-money valuation.
Amazon is committing $50bn, and SoftBank and Nvidia $30bn each. More financial investors are expected to join. Amazon’s money comes as $15bn now and $35bn “when certain conditions are met”. OpenAI also expanded its $38bn AWS contract by $100bn over eight years, committing to use about 2 GW of Amazon’s Trainium chips.
AWS becomes the exclusive third-party cloud distributor of OpenAI’s Frontier agent platform. From Nvidia, OpenAI secured 3 GW of dedicated inference capacity and 2 GW of training on Vera Rubin systems.
OpenAI says ChatGPT has >900m weekly active users and >50m consumer subscribers. Weekly Codex users have more than tripled this year to 1.6m.
Trump ordered every federal agency to stop using Anthropic.
“I am directing EVERY Federal Agency in the United States Government to IMMEDIATELY CEASE all use of Anthropic’s technology,” Trump wrote on Truth Social. The White House posted his statement that the US will never allow “A RADICAL LEFT, WOKE COMPANY” to dictate how its military fights. Agencies get six months to move to alternatives to Claude.
The Pentagon will cut its contract with Anthropic, worth up to $200m. Palantir, which runs its most sensitive military work on Claude, will likely need a deal with a rival.
Hegseth then directed the Pentagon to designate Anthropic a supply chain risk. Military contractors are barred, with immediate effect, from any commercial activity with it. Anthropic said it would challenge any designation in court, arguing under 10 USC 3252 the label can cover only the use of Claude on Department of War contracts.
Trump’s statement on Anthropic · The White House on X
Hegseth directs supply-chain risk designation · Pete Hegseth on X
Statement on the comments from Secretary of War Pete Hegseth · Anthropic
Trump moves to blacklist Anthropic’s Claude from government work · Axios
OpenAI agreed to deploy its models on the Pentagon’s classified network.
Sam Altman announced the deal the night the Pentagon moved to blacklist Anthropic. He said it includes OpenAI’s “prohibitions on domestic mass surveillance and human responsibility for the use of force”. OpenAI will deploy on cloud networks only, and asked the department “to offer these same terms to all AI companies”.
In a memo to staff on Thursday, Altman called the dispute “an issue for the whole industry” and said OpenAI’s red lines were mass surveillance and autonomous lethal weapons.
Agreement with the Department of War · Sam Altman on X
Sam Altman says OpenAI shares Anthropic’s red lines in Pentagon fight · Axios
Pentagon approves OpenAI safety red lines after dumping Anthropic · Axios
ALETH · OPEN ACCESS RESEARCH