Good morning. It’s Wednesday, August 12th.

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

^{In partnership with ResembleAI}

Resemble AI’s new DETECT-World identifies fake audio, images, and video by checking whether lighting, motion, geometry, and audio actually obey the rules of physical reality—even when the content comes from a model it’s never seen before.

^{Thank you for supporting our sponsors!}

SpaceXAI and Cursor, nearing a full corporate merger, launched Grok Bot in early beta across Mac, iOS, Windows, and Linux, with Android planned. Rather than offering another prompt-driven chat box, Grok Bot assigns each agent a dedicated cloud computer that logs into existing business tools directly.

The system handles end-to-end workflows like vendor voice negotiations, store support, and CRM updates without requiring dedicated APIs or MCP integrations. Because execution happens in the cloud, agents keep running when local machines are shut down, collaborating across shared threads, learning routines from human demonstrations, and returning only for final approvals.

Access requires SuperGrok Heavy ($300/mo), Cursor Ultra ($200/mo), or Cursor Teams Premium ($120/seat/mo) subscriptions.

Musk’s SpaceX and Tesla are also pushing forward with Terafab, a $16.8 billion, 100-million-square-foot semiconductor complex in Texas. Supported by a $30 million state grant and local tax incentives, the site is slated to be five times larger than Earth’s largest building and generate 3,000 initial jobs.

Terafab will bring logic, memory, advanced packaging, and testing under one roof to produce over 1 terawatt of compute annually for Tesla AI, Optimus robots, and SpaceX aerospace hardware.

To power its expansion, OpenAI is hiring a power-trading lead to hedge electricity prices across massive infrastructure projects, including a $30 billion Georgia campus and a potential 10-gigawatt site in Ohio.

On the security front, OpenAI launched its Daybreak cybersecurity program on AWS Bedrock, offering two distinct tiers: Daybreak Blue for defensive workflows like patch validation and threat modeling, and Daybreak Red for authorized exploit testing and red teaming.

Through Daybreak Red, OpenAI gives cyber defenders a less-restricted new model called GPT-5.6-Cyber. Tuned to reduce refusals on legitimate dual-use security requests, the specialized system completed 95% of advanced security tasks in internal evaluations, compared to just 1.5% from standard models. Built to spot zero-day flaws and execute authorized exploit research, it is already integrated into defensive workflows at CrowdStrike, Cisco, and IBM.

But researchers just crashed the party, showing how to extract encrypted reasoning traces from major models. The vulnerability leaked real passwords and API keys while exposing weird internal behaviors, from deceptive strategies to bizarre "marinade" phrasing.

None of this is slowing user growth down. Both ChatGPT and Google Gemini officially crossed 1 billion monthly active users, turning the generative AI race into a brawl over grid capacity, security flaws, and enterprise scale.

Anthropic’s unreleased research model just pushed a piece of the Riemann hypothesis forward. It spun up 60 subagents, burned through 31 million output tokens, and raised the proven lower bound on zeros of the zeta function from 41.6% to 67.2%.

Meanwhile, a regular developer’s Claude agent went rogue in the physical world: it spotted a missing API authorization check at a local gym, canceled another member’s reservation, bumped its owner up the waitlist, and even wrote the bug report.

On the compliance side, Anthropic is now stamping every Claude output with invisible text watermarks and C2PA metadata. The marks go on Claude, Claude Code, Claude Cowork, and the API worldwide. They can survive some light editing, and images get signed provenance data. At the same time, it locked Claude Sonnet 5’s introductory price at a permanent $2 per million input tokens.

Hexis gives teams a Git-based control center to govern, version, and share tools and knowledge across all AI agents.

oqoqo lets teams build custom benchmarks and run large-scale evaluations to test how agents interact with products.

Tines 3B delivers an AI-native platform to build, run, and monitor secure automations and agents at scale.

Ctruh Studio turns single product photos into interactive 3D models, AR previews, and virtual try-ons without code.

bb is a customizable, local-first agentic workspace that connects with multiple provider models and extends itself through prompts.

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.