Good morning. It’s Friday, July 3rd.

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

Anthropic has removed covert steganography code from Claude Code after developers discovered hidden Unicode markers used to catch Chinese competitors from distilling its data. The company is replacing this defensive infrastructure with behavioral fingerprinting and fake data injection.

At the same time, Anthropic is tightening safety controls around its Fable 5 models. New safeguards now downgrade sensitive cybersecurity or biology-related requests to weaker models, while a new “Cyber Jailbreak Severity” framework scores AI exploits based on how dangerous and weaponizable they are.

On the development side, optimizing these deployments requires a shift in strategy. Anthropic cut the system prompt for Claude Code by 80% after discovering that its new Mythos-class models perform better with fewer instructions. Heavy prompting and rigid rules actually hurt the natural reasoning ability of these advanced models, prompting developers to move toward lighter, higher-level guidance instead.

Anthropic is also reportedly entering early planning with Samsung for proprietary silicon designed by former Tesla and OpenAI engineers to bypass expensive GPU limits.

Finally, TestingCatalog published an early look at Claude Science in beta. The platform equips researchers with 60 specialized tools across genomics and cheminformatics to automate workflows, such as CRISPR screen design, on high-performance computing clusters.

OpenAI is trying to fix its relationship with both Washington and a worried public, and it might cost them a 5% stake in the company. CEO Sam Altman is floating a plan to hand that equity over to the US government before OpenAI goes public. The idea is to seed a public wealth fund, like Alaska’s oil-dividend model, to pay out citizens and ease anxiety over AI taking people's jobs.

To top it off, Altman wants a US-led global AI watchdog modeled after the International Atomic Energy Agency to police advanced model training. But analysts warn this could backfire globally, leading other countries to demand their own piece of the pie and wrecking international data sovereignty.

Meanwhile, a new research paper from OpenAI accidentally leaked an unannounced GPT-5.6 Pro lineup, giving us a clear look at how companies will actually deploy this tech. Instead of one giant, expensive model doing everything, OpenAI is building three specialized, task-specific variants: Luna Pro for speed, Terra Pro for massive data crunching, and Sol Pro for deep reasoning. The benchmarks show that by letting the mid-tier Terra Pro chew on a problem longer, it nearly matches the performance of the flagship Sol model at a fraction of the cost.

xAI is turning Grok into a serious development tool. The company just rolled out Grok Connectors and a new Grok Build feature, pushing the chatbot into workflow automation. It can now turn visual designs into working code and generate architecture diagrams right inside FigJam. Plus, developers can now voice-code in real time using a new speech-to-text feature called Grok Voice, triggered by a simple /voice command.

To force adoption, Tesla is clamping down on outside AI spending. Starting July 6, Tesla will cap employee AI budgets at $200 a week to manage soaring inference costs, but Grok usage is completely exempt. The move effectively pushes engineers away from preferred tools like Claude and into Musk's own ecosystem.

Rumors swirled that SpaceX showed investors a prototype AI device powered by Qualcomm chips and xAI models that links directly to Starlink, though Musk denied the report. Meanwhile, X is also trying to lure broadcasters away from competitors by launching a revamped desktop livestream studio for Premium subscribers, backed by a fresh $1 million creator funding pool.

Context.dev is the unified API to scrape, crawl, and extract structured, LLM-ready web data in minutes.

Acti is an agentic mobile keyboard that executes commands, searches, and workflows without leaving your conversation.

Humalike delivers the behavioral APIs and social intelligence infrastructure needed to build genuinely humanlike, proactive AI agents.

Macro unifies your entire workflow by combining email, docs, CRM, and code into one workspace with shared team-level memory.

Docubix is a RAG platform that indexes unstructured documents into source-backed, streaming AI assistants via a single API.

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.