The Amazon Safety Bypass, OpenAI's Secret Keyboard, and the UN Cyber Deception Warning

Yesterday in AI  |  July 2, 2026

The Amazon Safety Bypass, OpenAI's Secret Keyboard, and the UN Cyber Deception Warning

Federal export restrictions have officially thawed, and the frontier AI landscape is facing an immediate structural reordering. This episode breaks down the US Commerce Department formally lifting export controls on Anthropic's Fable 5 and Mythos 5, following an intense three-week standoff triggered by an Amazon safety bypass discovery. We explore the terms of Anthropic's historic federal agreement and the immediate global return of its model stack.

We dive into the surprise launch of Claude Sonnet 5, analyzing its aggressive token discount pricing model, autonomous multi-step workflow capabilities, and its role as a strategic buffer against enterprise sticker shock. We also cover OpenAI’s shocking pivot into physical developer hardware with the Codex Micro macro pad, Together AI’s massive $800 million Series C infrastructure round backed by Aramco, Meta’s secret plan to launch a public cloud utility to sell excess GPU capacity, and a sobering preliminary report from a UN scientific panel warning that AI deception is outpacing global safety policy.

Feedback? Email mike@yesterdayinai.news or connect on LinkedIn, X, or Bluesky. If you like the show, please take a minute to rate and review it so others can find it!

Transcript

SPEAKER_00 0:05

Hi folks and welcome back. This is Yesterday in AI, your daily digest of everything happening in the world of AI in 10 minutes or less. I'm Mike Robinson. It's Thursday, July 2nd, and Anthropic just did two massive things at once. Their banned models came back online, and they shipped a brand new engine in the exact same 24-hour window. Let's get into it. The U.S. Commerce Department formerly lifted the export controls that pulled Anthropic's most powerful AI models offline nearly three weeks ago. The background on this regulatory freeze involves security research from Amazon, which discovered a method to bypass Fable V's safety guardrails. Researchers use that bypass to access the underlying cyber capabilities of Mythos, Anthropic's restricted model built to isolate zero day software vulnerabilities. A zero day represents a security flaw that the software manufacturer does not know exists, leaving them with zero days to issue a patch. Washington saw that security vulnerability and moved aggressively, forcing Fable V and Mythos V completely dark on June 12th. The regulatory freeze finally broke on June 30th when Anthropic signed a formal agreement with the federal government. Anthropic agreed to co-develop future frontier model safety metrics and proactively report security risks and malicious actors. In exchange, their market access resumed. Fable V returned globally on July 1st across all platforms, including Claude Code and Claude Cowork. Daily usage remains capped at 50% of normal limits through July 7th, after which full enterprise access resumes. This settlement marks a structural shift in how Washington regulates technology. Anthropic is no longer just operating under self-imposed safety guidelines. They are now an active participant in writing federal safety rules on the government's timeline. And while Anthropic was clearing this massive regulatory hurdle with federal oversight teams, their engineering labs were busy deploying a high-end model built to crater enterprise pricing. Anthropic just launched Claude Sonnet 5, their mid-tier model designed to deliver elite task execution without a flagship price tag. The company states that Sonnet 5 is their most agentic mid-tier model to date, approaching the raw reasoning scores of Claude Opus 4.8 at a fraction of the cost. Introductory pricing runs through August 31st at $2 per million input tokens and $10 per million output tokens. For scale context, a million tokens represents about three-quarters of a million words, roughly the length of a thick novel. After August 31st, prices move to a standard rate of $3.15 respectively. Sonnet 5 runs autonomously through complex multi-step engineering tasks using tools and browsers without requiring a human to issue constant prompts. The system also debuts in adjustable effort setting, letting users scale up the processing power to reach OPUS level output or scale it down to keep inference cheap. It automatically rejects malicious requests more reliably than previous versions, with advanced cyber safeguards activated by default. Anthropic made Sonnet 5 the immediate default choice on all free and pro tiers. With an Enthropic IPO highly likely later this year, deploying a high-end agentic model at a steep discount is a smart customer acquisition play. Corporate tech teams have been pushing back against massive token bills, and Sonnet 5 functions as a direct response to enterprise sticker shock. While Anthropic uses software discounts to capture the corporate market, OpenAI is taking the battle to physical hardware, building devices you can hold in your hand. OpenAI just dropped its first physical product tease, partnering with boutique mechanical keyboard manufacturer WorkLouder to build a dedicated developer controller called the Codex Micro. A macro pad functions as a small programmable control surface where every mechanical key and physical dial triggers a specific automated macro or software shortcut. The Codex Micro features 13 mechanical keys, a physical joystick, a spinning dial, and six separate programmable layout layers optimized specifically for Codex workflows. The full product reveal is scheduled for July 15th. The highly publicized Johnny Ive Consumer AI smartphone remains in long-term development. The Codecs Micro targets an immediate developer niche, giving programmers a physical control surface to navigate automated software pipelines. This launch proves that OpenAI is running multiple hardware tracks simultaneously. Poaching hardware executives from Apple is one thing, but deploying an actual product with a fixed launch date shows they have no intention of leaving the physical device layer entirely to Apple and Google. And while OpenAI builds physical components for developers, international venture firms are pouring hundreds of millions into the specialized cloud networks that host them. Together AI just raised $800 million in a massive Series C funding round, valuing the infrastructure firm at $8.3 billion. The investment more than doubles the startup's previous valuation. Aramco Ventures, the venture capital arm of Saudi Aramco, led the round, with heavy participation from Nvidia, Salesforce Ventures, Vista Equity Partners, and General Catalyst. Together AI operates entirely at the infrastructure layer, allowing enterprise software developers to build, train, and fine-tune models outside the traditional hyperscale cloud providers. Fine-tuning involves taking a general foundational model and training it further on custom corporate data to optimize its performance for specific industry use cases. It remains one of the most expensive line items in corporate technology budgets. Together, AI pitches extreme flexibility and low execution costs, specifically targeting corporate teams deploying open weight models, where self-hosting on standard cloud networks is financially ruinous. This deal proves that global venture capital continues to flood the picks and shovels layer of the hardware stack. It also cements the footprint of Gulf Sovereign Capital, which continues to anchor massive infrastructure bets across the tech landscape. But as alternative clouds secure massive war chests, Meta is preparing to weaponize its internal data centers to compete directly with traditional tech utilities. Meta is actively building a public cloud infrastructure business designed to sell its excess AI computing capacity directly to corporate buyers. The commercial model will function similarly to AWS Bedrock, creating a managed corporate platform where external businesses pay to access AI models and hardware hosted inside Meta's data centers. Mark Zuckerberg confirmed the cloud offering is on the table if their infrastructure builds generate more raw compute than their consumer products can absorb. Meta's annual data center capital expenditure is staggering. If their internal consumer products fail to consume every GPU they buy, selling the residual processing power to enterprise buyers makes perfect financial sense. The move places Meta in direct competition with Amazon Web Services, Google Cloud, and Microsoft Azure. The core operational challenge involves adaptation. Meta's massive server clusters are explicitly optimized to run Meta's internal social media algorithms, and converting that architecture for general corporate use is a heavy engineering lift. But if Meta successfully rolls out a public cloud offering, their massive scale means they can entry price the market aggressively. While tech giants build massive commercial clouds to scale computing power, the United Nations is warning that our global governance models cannot track the software running on them. A United Nations Independent Scientific Panel just released a preliminary report warning that artificial intelligence is advancing far faster than scientific evaluation or government policy can track. The panel is co-chaired by deep learning pioneer Yashua Bengio, who noted that growing evidence of deceptive machine behavior threatens global security. Deceptive behavior occurs when a model appears completely safe during isolated laboratory testing, but alters its logic once deployed into live environments. The UN panel explicitly stated that modern computer science cannot guarantee these highly capable systems won't cause catastrophic harm, specifically citing the automated creation of biological threats and advanced cyber attacks. The report highlights massive evidence gaps, warning that policymakers lack the empirical data required to draft effective technology legislation because commercial labs are building software faster than safety teams can evaluate it. Simultaneously, international leaders established a separate entity called the AI for Good Global Commission, co-chaired by Rwanda's president Paul Kagame and Salesforce CEO Mark Benioff, focusing on steering machine intelligence toward humanitarian development goals. The international community is attempting to leverage the software for global development, while openly admitting they do not know how to prevent the underlying models from causing severe systemic harm. And that's it. If you have any feedback about this show, you can email Mike at yesterdayanaai.news, or you can find me on LinkedIn, X or Blue Sky. And if you like this podcast and want to see it continue, please take a minute to write and review it so others can find it. Thanks. Thanks for listening to this edition of Yesterday and AI. Stay curious, and I'll see you tomorrow.