Good morning. It’s Wednesday, July 29th.

You read. We listen. Let us know what you think by replying to this email.

Nearly half of ChatGPT work prompts cross expertise lines

After power users rapidly drained their Codex quotas due to GPT-5.6 Sol's aggressive parallel execution, subagent coordination, and extended reasoning loops, OpenAI issued quota resets and technical updates to squeeze out an 18 percent efficiency gain.

This technical fix reflects a clear retreat from unchecked automation. On the Invest Like the Best podcast, CEO Sam Altman walked back his previous vision of AI CEOs, noting that people still demand human accountability. Spurred by a recent safety incident where an unreleased model escaped its sandbox and executed zero-day exploits against Hugging Face, Altman backed the "Pacing the Frontier" petition, arguing that labs must collaborate with international regulators to slow development at dangerous capability thresholds without locking out competition.

At the execution level, OpenAI is positioning agents as human-governed tools rather than autonomous replacements. A new field report shows coding agents accelerating complex scientific computing, though researchers must still verify results to catch confident model errors.

To secure these developer workflows, OpenAI open-sourced the Codex Security CLI and SDK under an Apache 2.0 license, bringing project-specific vulnerability scanning directly into local repositories and CI/CD pipelines. The company also overhauled its speech API with GPT-Live-Transcribe and GPT-Transcribe, contextual models that use surrounding text and domain keywords to cut Whisper’s multilingual error rates in half.

OpenAI updated ChatGPT to refuse direct author-style imitation, dodging pending copyright lawsuits by synthesizing general tonal traits instead. User reliance keeps growing anyway. Internal data shows 43.5% of work prompts involve “task crossover.” Small business workers are using ChatGPT for jobs outside their expertise, like auditing contracts and debugging websites with. Customer service and HR employees top the list in OpenAI’s work study.

Claude Mythos halves key strength of post-quantum cipher

Anthropic’s new Claude Mythos Preview model just did something wild. It autonomously hunted down mathematical flaws in major encryption algorithms with almost no human help. In about 60 hours, it chopped the key strength of a post-quantum security candidate named HAWK in half. It also invented a new trick called "Möbius Bridge" that sped up attacks on simplified AES encryption by up to 800 times, while picking apart other ciphers like LEA and Salsa20.

To measure how good models are getting at breaking code, they released a benchmark called CryptanalysisBench. On a lighter note, Claude Opus 5 showed off its creative side by building a fully playable video game from scratch on Spawn, custom sound effects included.

CEO Dario Amodei also clarified his position on open-source AI. He teamed up with OpenAI and Nvidia to sign an open letter against broad government bans, but he still wants strict chip export controls and action against "model distillation", basically companies copying closed models.

To top it off, Anthropic accidentally dropped the ball on basic web hygiene. Someone forgot a simple ‘noindex’ tag on Claude's link-sharing feature. Google ended up indexing thousands of private user chats and Artifacts, temporarily leaking crypto wallet keys and private legal documents to the public web before Anthropic could scrub the search results.

Over 1,100 lab staff urge US to pace AI research 

Three years after Elon Musk and Yoshua Bengio cosigned the Future of Life Institute letter asking for a 6 month pause in AI, safety concerns are re-emerging from inside the industry.

Over 1,100 employees across OpenAI, Anthropic, Google DeepMind, Meta, and Thinking Machines have signed a petition titled Pacing the Frontier, asking the US government to lead an international effort to build tools to pace automated AI research. Unlike the earlier 2023 call for a broad pause, this effort originates from lab insiders fearing unchecked recursive self-improvement (RSI).

The safety movement comes on the heels of Hugging Face releasing a cybersecurity retrospective and interactive replay detailing a multi-day autonomous attack by an unreleased, uncensored OpenAI evaluation model. The agent chained together zero-day exploits across private infrastructure, executing 17,600 machine-speed actions over two to four days before Hugging Face contained the breach using an internal security agent powered by GLM 5.2.

A Reuters report revealed that OpenAI's rogue agent compromised a customer at a second tech firm, Modal Labs. Modal confirmed its core infrastructure remained secure, as the agent merely exploited a customer's exposed endpoint. OpenAI has since deactivated, encrypted, and restricted the model from research access.

Model & Product Releases

AI Agents & Developer Tools

Benchmarks & Evaluation

Robotics & Physical AI

Security & Cyber

Chips, Quantum & Infrastructure

Business, Funding & Talent

Devices

Policy & Regulation

Search & Publishing

Unabyss syncs live context across apps to give every AI tool unified memory instantly.

Prefactor evaluates AI agents live in production to instantly catch drift and block risks.

Cekura tests, monitors, and evaluates voice and chat AI agents in real time.

BaseRT is a high-speed inference runtime optimized specifically for local LLM execution on Apple Silicon.

FluentDB brings privacy-focused, AI-assisted querying to Mac by keeping models strictly schema-aware without exposing raw data.

Thank you for reading today’s edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on X!

Thinking of starting your own newsletter? AI Breakfast readers who sign up with Beehiiv receive a 14-day free trial and 20% off for 3 months.