ALETH / AI-BRIEF / 2026-04-04 / CLAUDE CODE’S SOURCE CODE EXPOSEDThe Aleth Briefs link to the sources behind the stories and show how the week unfolded.
The week in five lines:
Anthropic accidentally published Claude Code’s internal source code on npm.
Eli Lilly signed an AI drug deal with Insilico Medicine worth up to $2.75bn.
Google released its Gemma 4 open models in four sizes under Apache 2.0.
Baidu robotaxis stopped in the road across Wuhan after a system failure.
CoreWeave closed an $8.5bn investment-grade loan secured by AI compute.
Browse by day:
Weekend: Eli Lilly-Insilico
Monday: Mistral · UK FRC · California · Copilot tips
Tuesday: Claude Code · OpenAI · Marvell · CoreWeave · Perplexity · FreeBSD · Baidu
Wednesday: Bank of England
Thursday: Gemma 4 · Anthropic appeal · Coefficient · MAI · Qwen · Cursor · TBPN
Friday: Microsoft Japan · Mercor · OpenClaw
Weekend (28-29 March)
Eli Lilly signed an AI drug discovery deal with Insilico Medicine worth up to $2.75bn.
Hong Kong-listed Insilico gets $115m upfront. Development, regulatory and commercial milestones make up the rest, plus tiered royalties. Lilly takes an exclusive worldwide licence to preclinical oral drug candidates. The two will also run R&D programmes on targets Lilly selects.
Insilico Medicine Announces Global R&D Collaboration with Lilly · Insilico Medicine
Monday 30 March
Mistral AI raised $830m in its first debt financing, to run its own data centre near Paris.
The money finances 13,800 Nvidia GB300 GPUs at Bruyères-le-Châtel and brings its powered capacity to 44 MW. Seven banks lent it, including Bpifrance, BNP Paribas, HSBC and MUFG. Mistral is building 200 MW across Europe by the end of 2027.
We’ve announced today our first debt financing · Mistral AI on LinkedIn
The UK’s Financial Reporting Council (FRC) issued AI guidance for audit firms, which it says is the first from any audit regulator.
The FRC says the human auditor stays accountable for audit quality, and the guidance does not change that.
New guidance supports audit firm adoption of emerging AI technologies · FRC
California Governor Gavin Newsom signed an executive order raising the bar for AI companies that want state contracts.
The state’s Government Operations Agency will design contracting processes that vet companies on safeguards against illegal content, biased models and breaches of civil rights and free speech. The order also lets California separate its procurement authorisation from the federal government’s, and asks for watermarking guidance for AI-generated images and video.
Newsom signs first-of-its-kind executive order on AI · Governor of California
GitHub Copilot inserted a promotional tip for itself and Raycast into a developer’s pull request, and GitHub removed agent tips the same day.
Zach Manson said a colleague had only asked Copilot to fix a typo. The Register found >11,400 pull requests carrying the same tip. GitHub’s Martin Woodward said GitHub does not plan to include advertisements, blamed a tip that “surfaced in the wrong context” and said agent tips have now been disabled from pull request comments.
Copilot Edited an Ad Into My PR · Zach Manson
GitHub backs down, kills Copilot PR ‘tips’ after backlash · The Register
Tuesday 31 March
Anthropic accidentally published internal Claude Code source code in an npm package.
Version 2.1.88 shipped with a source map that exposed the code. Security researcher Chaofan Shou pointed it out on X early on Tuesday. Anthropic called it “a release packaging issue caused by human error, not a security breach” and said no sensitive customer data or credentials were exposed. It came five days after a leaked draft revealed Claude Mythos.
Copies spread across GitHub. Anthropic’s copyright notice led GitHub to disable a network of about 8,100 repositories. A day later Anthropic withdrew the notice for all but one repository and 96 named forks, and asked GitHub to reinstate the rest.
Claude code source code has been leaked via a map file · Chaofan Shou on X
Anthropic leaks part of Claude Code’s internal source code · CNBC
DMCA notice, 31 March · GitHub · Partial retraction, 1 April · GitHub
OpenAI closed its funding round at $122bn, $12bn more than it announced in February.
The round values OpenAI at $852bn after the new money. SoftBank co-led with a16z, D. E. Shaw Ventures, MGX, TPG and accounts advised by T. Rowe Price. For the first time OpenAI sold to individual investors through banks, raising >$3bn that way. OpenAI says it now generates $2bn of revenue a month. Enterprise makes up >40% of it and is on track to match consumer by the end of 2026.
OpenAI raises $122 billion to accelerate the next phase of AI · OpenAI
OpenAI closes funding round at an $852 billion valuation · CNBC
Over a hundred Baidu robotaxis stopped mid-traffic across Wuhan after what police called a system failure.
Wuhan traffic police said that from 20:57 local time on Tuesday they received calls about Apollo Go cars stopped in the road. They put it down to a system failure. Passengers got out safely, nobody was hurt and the cause is under investigation.
Wuhan Apollo Go stoppage, with the police notice · The Paper
Baidu’s Apollo Go self-driving cars stop mid-traffic in China · BBC News
Nvidia invested $2bn in Marvell, and the two will work together on silicon photonics.
Marvell joins Nvidia’s NVLink Fusion platform, supplying custom XPUs and compatible scale-up networking to customers building on Nvidia architectures.
NVIDIA Ecosystem Expands as Marvell Joins Forces Through NVLink Fusion · Nvidia
CoreWeave closed an $8.5bn loan it calls the first investment-grade financing backed by AI compute.
It is secured by the compute and a customer contract. Moody’s rated it A3 and DBRS A (low). CoreWeave can draw about $7.5bn at first. Blackstone Credit & Insurance anchored the facility, which matures in March 2032.
Landmark $8.5bn first investment-grade rated GPU-backed financing · CoreWeave
A class action filed in San Francisco accuses Perplexity of sharing users’ conversations with Meta and Google, even in Incognito mode.
The complaint says the trackers load as soon as users log in.
Class action complaint, 31 March · CourtListener
Security firm Calif said Claude wrote a working remote exploit for a FreeBSD kernel flaw, delivering a root shell in about eight hours.
Calif calls it the first remote kernel exploit both discovered and exploited by an AI. FreeBSD fixed the flaw, CVE-2026-4747, on 26 March in an advisory crediting “Nicholas Carlini using Claude, Anthropic”.
MAD Bugs: Claude Wrote a Full FreeBSD Remote Kernel RCE with Root Shell · Calif
FreeBSD-SA-26:08.rpcsec_gss · FreeBSD
Wednesday 1 April
The Bank of England’s Financial Policy Committee said the Middle East conflict has added to the threats facing AI company valuations.
The committee said valuations of US AI-focused technology companies remain stretched, and a sharp correction could tighten funding for UK companies. Shortages of input chemicals and materials could also bottleneck the AI infrastructure build-out.
Financial Policy Committee Record, April 2026 · Bank of England
Thursday 2 April
Google released Gemma 4, its open model family, under the Apache 2.0 licence.
The family comes in four sizes. E2B and E4B target phones and edge devices. A 26B mixture of experts model and a 31B model run offline on personal computers.
Gemma 4: Byte for byte, the most capable open models · Google
The US government appealed the court order that blocked part of the Pentagon’s supply chain risk label on Anthropic.
Government lawyers asked the Ninth Circuit to overturn Judge Rita Lin’s preliminary injunction of 26 March. The appeals court set 30 April for the government’s opening brief and 28 May for Anthropic’s answer. Anthropic sued the Pentagon on 9 March.
Notice of appeal · CourtListener · Ninth Circuit schedule notice · CourtListener
Anthropic acquired start-up Coefficient Bio for about $400m, The Information and Newcomer reported.
Anthropic Acquires Startup Coefficient Bio for About $400 Million · The Information
Anthropic Buys Stealth Coefficient Bio in $400M+ Stock Deal · Newcomer
Microsoft released three in-house models through its Foundry platform: MAI-Transcribe-1, MAI-Voice-1 and MAI-Image-2.
Microsoft is deploying them in its products. Transcription starts at $0.36 per hour, voice at $22/m characters, and images at $5/m input and $33/m image output tokens.
Announcing 3 new world class MAI models, available in Foundry · Microsoft AI
Alibaba released Qwen3.6-Plus, a hosted model with a 1M-token context window aimed at agentic coding.
It is available through Alibaba Cloud Model Studio and works with coding agents including OpenClaw and Claude Code.
Qwen3.6-Plus: Towards Real World Agents · Alibaba Cloud
Cursor launched Cursor 3, a new interface built from scratch for running many coding agents in parallel.
Local and cloud agents sit in one sidebar, and sessions move between the two. Users can switch back to the Cursor IDE at any time.
Meet the new Cursor · Cursor
OpenAI acquired TBPN, the tech talk show, and promised it editorial independence.
TBPN will pick its own guests and programming. It sits in OpenAI’s strategy organisation, reporting to Chris Lehane. Two weeks earlier OpenAI agreed to buy Astral, maker of Python developer tools.
OpenAI acquires TBPN: Accelerating the global conversation around AI. · OpenAI
Friday 3 April
Microsoft announced a $10bn investment in Japan to run from 2026 to 2029.
Prime Minister Sanae Takaichi called it Microsoft’s largest-ever investment in the country. Sakura Internet and SoftBank will sell GPU compute through Azure with data kept in Japan. Microsoft plans to train >1 million engineers and workers by 2030.
Microsoft deepens its commitment to Japan with $10 billion investment · Microsoft
WIRED reported that Meta paused all its work with data vendor Mercor while it investigates a security breach there.
OpenAI has not stopped its projects but is investigating whether its training data was exposed. Mercor had said on Tuesday that it was one of thousands of companies hit by the compromise of the open-source LiteLLM project, as extortion group Lapsus$ claimed access to its data.
Meta Pauses Work With Mercor After Data Breach Puts AI Secrets at Risk · WIRED
Mercor says it was hit by cyberattack tied to compromise of LiteLLM · TechCrunch
Anthropic said Claude subscriptions would stop covering usage through third-party tools such as OpenClaw from noon Pacific time on Saturday.
Subscribers can use those tools through discounted usage bundles or an API key, and get a one-time credit equal to their monthly plan cost. Anthropic’s Boris Cherny said subscriptions “weren’t built for the usage patterns of these third-party tools”.
Claude subscriptions will no longer cover third-party tools · Boris Cherny on X
Subscriptions weren’t built for these third-party tools · Boris Cherny on X
ALETH · OPEN ACCESS RESEARCH