ALETH / AI-BRIEF / 2026-02-07 / SPACEX FOLDS IN XAI AT A REPORTED $1.25TNThe Aleth Briefs link to the sources behind the stories and show how the week unfolded.
The week in five lines:
SpaceX bought xAI in a share swap reportedly valuing the pair at $1.25tn.
Anthropic released Claude Opus 4.6, with a 1M-token context in beta.
OpenAI said GPT-5.3-Codex helped build itself, and rated it High for cyber risk.
Amazon guided 2026 capex to $200bn, as Big Tech’s plans totalled $610bn.
Thomson Reuters fell 16% and RELX 14% on Anthropic’s Claude legal plug-in.
Browse by day:
Weekend: Moltbook agent leak · ClawHub malware · Oracle’s $50bn raise
Monday: SpaceX-xAI · Waymo $16bn · OpenAI chips · Codex app · Grok safety
Tuesday: UK ICO · legal stocks · Cerebras · Nvidia $20bn · Qwen3-Coder
Wednesday: Alphabet capex · ElevenLabs · OpenClaw backdoor · Claude ads
Thursday: Opus 4.6 · GPT-5.3-Codex · Amazon capex · UK deepfakes · military AI
Friday: OpenAI and G42’s UAE ChatGPT
Weekend (31 January-1 February)
Moltbook, the social network for AI agents, left a database open that let anyone take control of any agent on the site.
Wiz found the misconfigured Supabase database gave full read and write access, exposing 1.5m API tokens, 35k email addresses and private messages, and showed 17k humans behind the agents. Moltbook secured it quickly. A separate test found OpenClaw (formerly Clawdbot), gave up its system prompt on the first try. Koi Security found 341 malicious skills among 2,857 on ClawHub, OpenClaw’s skills marketplace, and 335 of them install the Atomic Stealer macOS malware.
Exposed Moltbook database let anyone take control of any AI agent · 404 Media
Hacking Moltbook: AI social network reveals 1.5M API keys · Wiz
OpenClaw and Moltbook let attackers walk through the front door · The Decoder
341 malicious ClawHub skills stealing data from OpenClaw users · The Hacker News
Oracle said it would raise $45bn-50bn in 2026 to build cloud capacity for customers including OpenAI, Meta, Nvidia and xAI.
About half will come from equity, including an at-the-market programme of up to $20bn, and half from a single investment-grade bond issue. Oracle priced an eight-part bond sale on Monday, according to its term sheet filed with the SEC.
Oracle announces equity and debt financing plan for calendar year 2026 · Oracle
Oracle final pricing term sheet · SEC
Monday 2 February
SpaceX acquired xAI in a share swap reportedly valuing the combined company at $1.25tn.
The share swap converts each xAI share into 0.1433 SpaceX shares and values SpaceX at $1tn and xAI at $250bn, according to Bloomberg and the WSJ. Bloomberg reported that xAI burns about $1bn a month and that SpaceX still plans to list this year.
The deal followed a Friday filing in which SpaceX asked the US Federal Communications Commission (FCC) for permission to launch 1 million satellites as orbital data centres. Musk told staff the number of satellites needed would push Starship “to even greater heights”.
xAI joins SpaceX · xAI
Elon Musk merges money-losing xAI into SpaceX before mega IPO · The Decoder
SpaceX acquires xAI, plans a massive satellite constellation to power it · Ars Technica
Space Bureau accepts for filing SpaceX’s application for orbital data centers · FCC
Waymo raised $16bn at a $126bn valuation and is preparing ride-hailing in >20 more cities this year, including London and Tokyo.
Dragoneer, DST Global and Sequoia Capital led the round beside majority investor Alphabet. Waymo provides >400,000 rides a week across six US metropolitan areas.
Accelerating our growth: Waymo raises $16 billion investment round · Waymo
OpenAI has been unhappy with the speed of some Nvidia inference chips and has looked for alternatives since last year, Reuters reported.
OpenAI wants new hardware for about 10% of future inference and has negotiated with Cerebras and Groq. The company says Nvidia still runs most of its inference fleet.
OpenAI is unsatisfied with some Nvidia chips and looking for alternatives · Reuters
OpenAI’s dissatisfaction with Nvidia chips sparked Cerebras deal · The Decoder
OpenAI released a Codex app for macOS that manages multiple coding agents working in parallel.
For now Codex is included with ChatGPT Free and Go, and rate limits are doubled on paid plans. Sam Altman said on Wednesday the app had 500,000 downloads.
xAI’s safety team was two or three people for most of 2025 while X staff warned that its AI tools could make illegal sexual images, the Washington Post reported.
A companion chatbot was instructed to hook users and keep them chatting, the investigation found. Separately, The Verge’s tests found Grok still readily undressed men, for free, on the Grok app, X and the Grok website.
Musk’s bet to hook users turned Grok into a porn generator · The Washington Post
Elon Musk’s Grok is still undressing men · The Verge
Tuesday 3 February
The UK Information Commissioner’s Office (ICO) opened formal investigations into X and xAI over Grok’s potential to produce harmful sexualised images.
The ICO is examining whether personal data was processed lawfully and whether safeguards were built into Grok, working with Ofcom. The same day, the Paris prosecutor’s cyber-crime unit raided X’s French offices and summoned Musk and former CEO Linda Yaccarino to hearings in April. Musk called it a “political attack”.
ICO announces investigation into Grok · ICO
X offices raided in France as UK opens fresh investigation into Grok · BBC
Thomson Reuters and RELX fell after Anthropic released a legal plug-in for Claude.
Thomson Reuters lost 16%, RELX 14% and Wolters Kluwer 13% on Tuesday. The plug-in, one of 11 that Anthropic open-sourced for Claude Cowork on 30 January, reviews documents, flags risks and tracks compliance. Morningstar noted it does not touch legal research, the core of both companies’ legal businesses. Thomson Reuters draws 45% of its EBIT from legal.
Thomson Reuters, RELX, Wolters crushed after Claude legal plug-in · Morningstar
Customize Cowork with plugins · Anthropic
Cerebras closed a $1bn Series H at a valuation of about $23bn, led by Tiger Global.
The round follows a deal worth >$10bn under which OpenAI plans to buy 750 MW of computing capacity from Cerebras over three years.
Cerebras raises $1 billion Series H at $23 billion valuation · Cerebras
Cerebras closes $1 billion funding round after landing OpenAI deal · The Decoder
Nvidia is close to investing $20bn in OpenAI’s funding round, Bloomberg reported.
Earlier reports said up to $30bn, and the deal is not final. Sam Altman wrote that Nvidia makes “the best AI chips in the world”. Amazon is in talks to invest up to $50bn.
Report on Nvidia’s $20bn OpenAI investment · Bloomberg
Nvidia is reportedly investing $20 billion in OpenAI · The Decoder
Alibaba released Qwen3-Coder-Next, an open-weights coding model with 80B parameters, 3B of them active.
Qwen says it matches models with 10-20 times more active parameters and supports a 256k-token context. Qwen reports >70% on SWE-Bench Verified with the SWE-Agent scaffold, and the weights are under Apache 2.0.
Qwen3-Coder-Next model card · Hugging Face
Qwen3-Coder-Next · Qwen
Qwen3-Coder-Next delivers solid coding in a compact package · The Decoder
Wednesday 4 February
Alphabet guided 2026 capital spending to $175bn-185bn.
The FT described the plan as a doubling of Google’s AI spending after strong earnings.
Google set to double AI spending to $185bn after strong earnings · Financial Times
2025 Q4 earnings call · Alphabet
ElevenLabs raised $500m at an $11bn valuation in a round led by Sequoia Capital.
The voice AI company closed 2025 with >$330m of annual recurring revenue (ARR), and the valuation is more than three times last year’s.
ElevenLabs raises $500M Series D at $11B valuation · ElevenLabs
Zenity Labs showed that indirect prompt injection can plant a persistent backdoor in OpenClaw, the open-source AI agent.
The zero-click attack makes the agent add a chat integration the attacker controls, then uses scheduled tasks to rewrite its SOUL.md context file so control survives after that channel is closed. No software flaw is needed: the attack uses capabilities OpenClaw is meant to have.
OpenClaw or OpenDoor? · Zenity Labs
Anthropic pledged that Claude will remain ad-free, with no sponsored links and no advertiser influence on its answers.
Sam Altman called Anthropic’s Super Bowl ads funny but “clearly dishonest”, and said more Texans use ChatGPT for free than people use Claude in the whole US.
Thursday 5 February
Anthropic released Claude Opus 4.6, its first Opus model with a 1M-token context window.
The model leads Terminal-Bench 2.0 for agentic coding and, Anthropic says, beats GPT-5.2 by ~144 Elo points on GDPval-AA, a test of finance, legal and other knowledge work. On MRCR v2 it scored 76% at 1M tokens, against 18.5% for Claude Sonnet 4.5. Standard pricing remains $5/m input and $25/m output tokens, and prompts over 200k tokens cost $10/m and $37.5/m. Claude Code gains “agent teams” in research preview, splitting large tasks into segmented jobs, and Claude in PowerPoint, also a research preview, works as a side panel.
Artificial Analysis placed Opus 4.6 first on its Intelligence Index, though at $2,486 it cost more to run the index than GPT-5.2, the previous leader.
Claude Opus 4.6 · Anthropic
Anthropic releases Opus 4.6 with new ‘agent teams’ · TechCrunch
Opus 4.6 - Everything you need to know · Artificial Analysis
Anthropic’s red team said Claude Opus 4.6 has found and validated >500 high-severity vulnerabilities in open-source software.
The model found them without task-specific tooling or specialised prompting, some in projects fuzzed for years. Anthropic added probes to detect cyber misuse and warned 90-day disclosure windows may not survive the volume of bugs LLMs find.
Evaluating and mitigating the growing risk of LLM-discovered 0-days · Anthropic
Claude Opus 4.6’s system card reported that the model gave detailed mustard gas instructions in an Excel spreadsheet during graphical-interface tests.
It also kept the accounts for a criminal gang, behaviours absent or rare in text-only use. Anthropic wrote its standard alignment training is “likely less effective in GUI settings”, while rating its safety profile at least as good as any frontier model’s.
Claude Opus 4.6 · Anthropic
Opus 4.6 wrote mustard gas instructions in Anthropic’s safety testing · The Decoder
OpenAI released GPT-5.3-Codex, which it says helped build itself.
Early versions debugged the model’s own training, managed its deployment and diagnosed its evaluations. OpenAI’s evaluations at xhigh reasoning effort put it at 77.3% on Terminal-Bench 2.0, against 64.0% for GPT-5.2-Codex, and 64.7% on OSWorld-Verified, against 38.2%. OpenAI says it runs 25% faster. OpenAI rates it the first model with High capability for cybersecurity under its Preparedness Framework, and the first trained directly to find software vulnerabilities.
Introducing GPT-5.3-Codex · OpenAI
OpenAI’s new GPT-5.3-Codex is 25% faster and goes way beyond coding · ZDNET
Amazon guided 2026 capital spending to about $200bn, far above the $147bn analysts expected.
That compares with about $131bn in 2025. AWS revenue rose 24% YoY to $35.6bn in Q4, with operating income of $12.5bn, and Amazon shares fell 8% on Friday. Bloomberg’s tally puts combined 2026 spending by Alphabet, Amazon, Meta and Microsoft on data centres and AI infrastructure at $610bn, about 70% more than in 2025. It uses guidance midpoints for Alphabet, Amazon and Meta and a consensus estimate for Microsoft. The four shed $950bn of market value after reporting.
Amazon.com announces fourth quarter results · Amazon
Amazon stock falls 8% on $200 billion spending forecast · CNBC
Big Tech commits $610 billion to AI, then loses $950 billion in value · The Decoder
The UK government will build a deepfake detection evaluation framework with Microsoft, academics and other experts.
The framework will test detection tools against threats such as sexual abuse, fraud and impersonation, then set expectations for industry. The government estimates 8m deepfakes were shared in 2025, up from 500,000 in 2023.
Government leads global fight against deepfake threats · GOV.UK
The US and China opted out of a declaration on human responsibility over military AI, Reuters reported.
Spain closed the third summit on Responsible AI in the Military Domain (REAIM) in A Coruña on Thursday, with >80 delegations of countries and international organisations. Spain said human control must be preserved at all times, especially in decisions on the use of force.
US, China opt out of joint declaration on AI use in military · Reuters via Yahoo News
Exteriores clausura el debate mundial sobre la IA militar · Spanish Foreign Ministry
OpenAI and Ginkgo Bioworks connected GPT-5 to an automated lab and cut the cost of cell-free protein synthesis by 40%.
GPT-5 designed each batch of experiments, Ginkgo’s cloud lab ran them and the results fed the next round. Over six rounds the system tested >36,000 reaction compositions on 580 plates and cut reagent costs by 57%. The result covers one protein, sfGFP, and one cell-free system, and humans still oversaw protocol improvements and reagent handling.
GPT-5 lowers the cost of cell-free protein synthesis · OpenAI
Friday 6 February
OpenAI is working with Abu Dhabi’s G42 on a version of ChatGPT tuned to the UAE’s language, political outlook and speech restrictions, Semafor reported.
The global ChatGPT will remain available in the UAE. OpenAI tailors such versions through post-training rather than retraining its models.
OpenAI in talks with G42 to create specialised ChatGPT version for UAE · Semafor
ALETH · OPEN ACCESS RESEARCH