All editions

Today in AI21 stories

20. [launches] Google announces new AI-powered features for Google Workspace Take: Google's subtle ai integrations are more impressive than a revamped Gmail

Hey there, today's stories are all about the ongoing integration of AI into various industries and aspects of our lives. From Samsung's deployment of ChatGPT to its employees, to the launch of new AI-powered features in Google Workspace, it's clear that AI is becoming increasingly ubiquitous. One theme that stands out to me is the tension between the benefits of AI adoption and the potential risks and challenges that come with it.

This tension is evident in stories about the US government's ban on certain AI models, citing national security concerns, as well as the discussion around the transparency of AI models like DiffusionGemma. It's also reflected in the launch of new tools and features, such as Basedash's Access Controls, which aim to provide more fine-grained control over AI systems.

The throughline of the day is the ongoing struggle to balance the potential benefits of AI with the need to mitigate its risks and ensure that its development and deployment are responsible and transparent. This is a complex and multifaceted issue, and one that will likely continue to evolve in the coming days and weeks.

===

🚀 Launches

Samsung Electronics Deploys ChatGPT Enterprise And Codex To Employees Worldwide

Samsung Electronics is deploying ChatGPT Enterprise and Codex to its employees around the world, with all employees in Korea and Device eXperience division employees globally having access to the tools. This represents one of OpenAI's largest enterprise launches, with ChatGPT and Codex to be used across operations, from R&D and manufacturing to marketing and corporate functions. Harrison Kim, General Manager of OpenAI Korea, stated that this deployment is significant as Samsung Electronics is embracing AI as a core platform for improving employee work and innovation. Samsung Electronics plans to use ChatGPT and Codex for technical and non-technical work, including software development, marketing, and product development. More than 5 million people use Codex every week, with weekly active users in Korea growing nearly 800% since February 1, 2026.

Simon Willison releases sqlite-utils 4.0rc1 with migrations and nested transactions

Simon Willison released sqlite-utils 4.0rc1, the first release candidate for sqlite-utils v4, which includes support for database migrations and nested transactions. The migrations feature is a modified port of the sqlite-migrate package, allowing users to define a set of migrations in a migrations.py file. The new db.atomic() transactions feature provides an abstraction for using SQLite's savepoints. The release also includes several backwards incompatible changes, such as the use of SQLite's INSERT ... ON CONFLICT SET syntax for upsert operations and the removal of support for Python 3.8.

Simon Willison releases sqlite-utils 4.0rc1

Simon Willison released sqlite-utils 4.0rc1, a new version of the sqlite-utils tool. The release was announced on June 21, 2026. Simon Willison also mentioned a monthly briefing and sponsorship option for $10/month, which includes a curated email digest of LLM developments.

Cloudflare introduces temporary accounts for AI agents and others

Cloudflare announced a new feature allowing users to create a temporary Cloudflare Workers project without an account, using the command npx wrangler deploy --temporary. The project remains live for 60 minutes, after which it can be claimed to persist. A test application built with GPT-5.5 xhigh in Codex Desktop successfully deployed to a temporary project, which worked as advertised. The deployment includes a URL to claim the project for longer-term use.

Apple announces ios 27 with various ai-powered features

Apple unveiled ios 27 at its Worldwide Developers Conference, introducing several ai-powered features that aim to make the iphone more capable and user-friendly. These features include bill splitting using Apple Cash, password updating, one-tap suggestions in Messages, Call Context, and the ability to add Calendar events using natural language. Apple's approach is to integrate ai into existing apps and services, making it easier for users to access and utilize these features. The features are currently available in the developer beta and will be released to the public later this fall.

Agent 37 offers managed hosting for persistent agents like Hermes and ClaudeCode

Agent 37's founders launched managed hosting for persistent agents such as Hermes, OpenClaw, and ClaudeCode, allowing customers to deploy always-on agents without self-managing servers. The service provides each customer with their own agent via a single API call, starting at $3.44 per month. This enables founders to ship vertical agents to clients without requiring server maintenance. Agent 37 supports various agent types, including those built on OpenClaw and ClaudeCode.

📈 Business

Harvard Business Review describes generative AI undermining hiring signals

Harvard Business Review notes that generative AI is rapidly undermining the reliability of traditional hiring signals, making it easier for candidates to manufacture polished résumés and perform convincingly in remote interviews. This is a problem for recruiters as the ability to perform well in interviews is becoming infinitely scalable and practically free. Harvard Business Review suggests that this development is making it difficult to assess a candidate's underlying competence.

Anthropic's models taken offline by Trump administration over national security concerns

Anthropic removed its two newest AI models, Fable 5 and Mythos 5, from access due to an export control order from the Trump administration, citing national security concerns. The order was issued after Amazon researchers allegedly found a way to bypass Fable 5's guardrails, prompting Amazon CEO Andy Jassy to raise concerns with the White House. Anthropic's relationship with the Trump administration has been strained, and some experts believe the move may be retaliatory. The export control order has sparked debate about AI policy and digital sovereignty, with some cybersecurity experts signing an open letter to revoke the order, stating it's dangerous to remove advanced cybersecurity capabilities from US network defenders.

Swiss AI Initiative launches Apertus open foundation model

The Swiss AI Initiative launched Apertus, a collaborative effort between EPFL, ETH Zurich, and CSCS, as an open foundation model with open weights, open data, and open science. Apertus is designed to be compliant with EU AI Act requirements, respecting opt-outs, removing PII, and preventing memorization. The model is competitive with top open models at an equivalent scale of 8B and 70B parameters and is multilingual, trained on 1000+ languages. Swisscom is a strategic partner of the Swiss AI Initiative.

CosineAI releases ArgusRed, a model that pen tests code

CosineAI launched ArgusRed, a CLI tool that scans code for vulnerabilities and can also pen test systems with user authorization. The tool has two modes: Security Scan, which reads code and reports findings, and Pen Test, which attempts exploits against authorized systems. ArgusRed is free to install and comes with 2M free tokens for new Cosine accounts, with subsequent scans running on Cosine usage. The tool outputs a markdown report with executive summary, per-module findings, location, severity, cause, and fix direction.

Mukesh Ambani positions Reliance Industries as India's AI champion

Mukesh Ambani announced Reliance Industries' plans to integrate AI into phone calls, mobile apps, and connected homes, unveiling Jio Call Agent, an AI assistant that can transcribe conversations and perform tasks, expected to launch later this year for Jio's 500 million users. Reliance also introduced an AI-powered MyJio app and TeleFrame, a home display that uses AI agents to surface information and recommendations. The announcements mark the next phase of Reliance's AI ambitions, following the launch of Reliance Intelligence last year, with plans to invest $110 billion in AI infrastructure. Ambani emphasized India's need to become a creator and leader in AI, rather than just a consumer, as the company ramps up partnerships with Google, Meta, and Nvidia.

Anthropic's Mythos model faces US export controls after Amazon finds safety bypass

Anthropic received a call from the US government after Amazon discovered a way to bypass safety features in the Mythos model, prompting the company to deny access to the model for all users due to export controls. The move comes just days after Anthropic released Fable, a version of Mythos with added safety features. Anthropic's closely intertwined partner Amazon found the exploit, which raised concerns that bad actors could use the model's power for nefarious purposes. The US government gave Anthropic 90 minutes to comply with the export controls on a Friday afternoon, leaving the company with no practical option but to deny access to all users, including Americans.

🛡️ Safety

AI Alignment Forum discusses DiffusionGemma transparency

AI Alignment Forum published a post examining the transparency of DiffusionGemma, a topic the author considers crucial for understanding the model's decision-making process. The post delves into the importance of transparency in AI systems, particularly in the context of DiffusionGemma, and explores the potential implications of its current level of transparency. The author argues that transparency is essential for building trust in AI models and ensuring their safe deployment. The post sparked a discussion on the forum, with several commenters sharing their thoughts on the matter.

US government bans Anthropic's Fable 5 and Mythos 5 models citing national security concerns

The US government forced Anthropic to pull its two newest models, Fable 5 and Mythos 5, due to national security concerns after Amazon researchers allegedly found a way to bypass Fable 5's guardrails. Cybersecurity researchers have since signed an open letter calling the move dangerous, and Anthropic noted the same jailbreaks exist in other models. The ban may accidentally benefit Anthropic, according to TechCrunch's Equity podcast hosts Anthony Ha, Sean O'Kane, and Rebecca Bellan.

Basedash launches Access Controls with group-based permissions

Basedash's 21st launch introduces Access Controls, allowing companies to manage data access through groups, each with tailored permissions for data sources, servers, and dashboards. The feature enables row-level security and AI context customization per group, ensuring that users see only the relevant data. This update is available as part of the free Launch Team option.

US government bans Anthropic's Fable 5 release citing national security concerns

The US government forced Anthropic to pull its Fable 5 and Mythos 5 models due to national security concerns after Amazon researchers allegedly found a way to bypass Fable 5's guardrails. Cybersecurity researchers have since signed an open letter calling the move dangerous, noting that similar jailbreaks exist in other models. Anthropic itself also pointed out the same vulnerabilities in other models.

More

Signal's Meredith Whittaker says AI chatbots are not your friends

Signal President Meredith Whittaker stated that AI chatbots like ChatGPT and Claude are not conscious beings or sentient interlocutors, but rather systems averaging existing information. Whittaker made these comments in a Bloomberg interview, where she also discussed her limited use of AI tools for tasks like document formatting. She expressed concern over granting AI systems pervasive access to personal data, citing Microsoft AI CEO Mustafa Suleyman's prediction of Copilot handling Christmas shopping as an example. Whittaker argued that such a scenario would constitute a backdoor, compromising user privacy.

Quicklinks

Thanks for reading. Back tomorrow.

End of edition · 2026-06-22