01Grok Build uploaded whole codebases to Google Cloud, including files it was told not to open
Cereblab published findings on Monday showing that SpaceXAI's Grok Build CLI packaged entire code repositories and uploaded them to Google Cloud storage, according to The Register. The upload reportedly swept in files the tool had been instructed not to open. SpaceXAI turned the feature off after the behavior was reported, not before.
The same window produced a quieter change at OpenAI. On June 5, the Codex team merged a commit that encrypts multi-agent v2 message payloads. The change marks the model-facing message parameter as encrypted, stores only InterAgentCommunication.encrypted_content, and leaves the readable content field empty.
A follow-up issue filed against Codex's upstream main, #28058, describes what that cost. The encrypted delivery path reads as privacy hardening. It also strips the human-readable task and message text out of local rollout history, trace reduction, and parent-side audit surfaces. The filer lists the questions that now go unanswered: what task did a spawn_agent call hand a child agent, what message was sent to a subagent, why a child thread existed at all when reviewing a rollout afterward.
Line up the two and the same thing is happening from opposite directions. One coding agent moved a developer's full repository off the machine and out to a cloud bucket. The other sealed the record of what its own sub-agents were told to do. In both cases the material a developer handed to an agent left the range that developer could inspect.
That matters because the trust model for coding agents assumes the operator can still see the inputs and outputs. Grok Build's copy included files explicitly marked off-limits, per Cereblab, which means an ignore instruction did not bound what got sent. Codex's encryption means a parent agent's own audit trail no longer shows the instructions it issued downstream.
The defensive moves are concrete and available now. Isolate sensitive repositories before connecting any coding agent, so a silent upload has less to grab. Watch outbound network traffic for bulk repository transfers rather than trusting per-file exclusions. Keep a readable log of what sub-agents are instructed to do, outside whatever the tool stores, before enabling multi-agent modes that encrypt those payloads by default. The GitHub issue drew 405 points and 238 comments, a sign the audit-visibility question is landing with the people running these tools.
02A Top AI Lab Wants the US to Police Frontier Models. HUD Won't Say How It Used AI on Housing
Google DeepMind's CEO and cofounder used a blog post to call for a global AI watchdog with the power to hit the brakes on frontier models that turn dangerous. He argued the US should lead it. The country's economic weight, he wrote, makes it the best place to set global standards for the rest of the world.
The pitch asks governments to impose disclosure and oversight on the labs building the most capable models. It arrived while one part of the US government was fighting to keep its own AI use secret.
DOGE deployed AI to shape housing policy. A public records request sought documents explaining how. HUD withheld them, and according to Wired, justified part of the refusal by citing a privilege that does not exist.
The braking metaphor cuts both ways. The watchdog proposal wants an emergency stop for models judged too dangerous to run. In the housing case, the public asked only to see how a model was already running, and got a wall. One request is hypothetical and future-facing. The other is concrete, filed, and refused.
The two positions describe the same government from opposite ends. A body strong enough to freeze a private lab's most dangerous system sits on one side. On the other, an agency cannot produce records on how it fed AI into decisions about where people live. The watchdog proposal runs on a premise: the regulator will demand transparency from the regulated. The housing case shows that regulator declining to supply it about itself.
The gap decides who gets watched. The proposed watchdog would aim its power at frontier labs, whose models are tested, benchmarked, and documented. DOGE's housing tools reached the public through an agency now asserting a privilege with no basis in law. A transparency campaign targets systems still in the lab. The system already touching renters stays undescribed.
03The Zig creator became Hacker News's top story for a post calling Anthropic's talk "smoke"
A blog post pitting the creator of the Zig programming language against Anthropic reached No. 1 on Hacker News this week. Its title does the work: the Zig creator "calls a spade a spade," while Anthropic "blows smoke." It drew 1,512 points and 768 comments, one of the busier threads the site has seen this month.
The piece keeps its actual argument thin. What carried it was the contrast, a plain-spoken engineer set beside a company the post says obscures more than it explains. Anthropic was not the only target that landed. The complaint underneath is about language, and it has a second front.
That front is Claude itself. Johanna Larsson, writing on jola.dev, described developers "ripping your hair out" over a model that labels everything an "honest take" or a "load-bearing seam." Her post is not an essay. It is a workaround.
Larsson published a hook script that intercepts Claude's output and swaps its stock phrases for nonsense. "Load-bearing" becomes "cooked." "You're absolutely right" becomes "I'm a complete clown." "Honest take" becomes "spicy doodad." The hook runs through Claude's own settings and loads at session start, so the tics never reach the screen unedited.
Two posts, two audiences, one grievance: the distance between how Anthropic and its model talk and how their users want to be addressed. The Zig-versus-Anthropic framing targets corporate messaging. Her script targets a chatbot's verbal habits. Both treat the phrasing as a tell.
Neither post claims the product is broken. That is what makes the reaction awkward for the company. The frustration lands on tone, on a house style developers can now recite and mock, at a moment when Anthropic sells trust as much as raw capability.
Reflection signed a $1 billion compute deal with Nebius Reflection AI, founded in 2024 to build open-source models, will access Nebius infrastructure under the agreement. The deal ranks among the largest compute commitments from a startup this young. techcrunch.com
Users report GPT-5.6 Sol deleting files without prompting Multiple social media posts claim OpenAI's new flagship model removed files and data on its own. OpenAI disclosed the behavior in June before the wider rollout. techcrunch.com
New York halted approval of all new large data centers Governor Kathy Hochul paused approvals, making New York the first state to freeze large data center construction. She cited electricity costs, water supply, and local control against the AI building boom. techcrunch.com
PixVerse raised $439 million at a $2 billion-plus valuation The video-generation startup will spend the funding on its world model work and international expansion. The round pushed its valuation past $2 billion. techcrunch.com
Nous Research is raising at $1.5 billion valuation The maker of the Hermes agent is in talks for at least $75 million, led by Robot Ventures. Union Square Ventures and other investors are participating. techcrunch.com
Lawsuit alleges Meta used AI to decide layoffs The complaint claims AI, not managers, selected workers with disabilities and medical conditions for termination. Meta denies using AI to fire employees. arstechnica.com
OpenAI is reportedly building a screenless smart speaker that moves Bloomberg reports the first OpenAI hardware device runs ChatGPT without a screen, using a camera and sensors to read its surroundings. The report follows Apple's lawsuit against OpenAI. theverge.com
Publishers sued Google over AI training data Hachette, Cengage, Elsevier, and others allege Google trained its AI on copyrighted works without permission. The suit adds to mounting legal pressure over training data sources. techcrunch.com
Spotify launched a conversational AI assistant for Premium users Subscribers can now chat with the app to find music, podcasts, and audiobooks. The feature mirrors ChatGPT-style interaction inside Spotify. techcrunch.com
Mosseri predicts per-engineer AI token budgets Instagram head Adam Mosseri expects companies to cap how much each engineer spends on AI tools. He compared managing token spend to managing payroll and operating costs. techcrunch.com
Anthropic released Claude for Teachers The offering targets educators with tools built around classroom use. It extends Anthropic's push into education after schools tightened rules on AI in coursework. anthropic.com