In partnership with
Good morning. It’s Friday, May 29th.
The vibe this week is basically "here's your agent's company credit card" and "by the way nothing is secure" published on the same day by different parts of the same industry.
Nobody's stopping, nobody's slowing down, and the closest thing to a warning label is a Cisco report that quietly dropped saying no frontier model can survive a multi-turn attack (which approximately zero people building agent infrastructure will read before shipping.)
-Jeff
AI Breakfast
You read. We listen. Let us know what you think by replying to this email.
Your prompts are leaving out 80% of what you're thinking.
When you type a prompt, you summarize. When you speak one, you explain. Wispr Flow captures your full reasoning — constraints, edge cases, examples, tone — and turns it into clean, structured text you paste into ChatGPT, Claude, or any AI tool. The difference shows up immediately. More context in, fewer follow-ups out.
89% of messages sent with zero edits. Used by teams at OpenAI, Vercel, and Clay. Try Wispr Flow free — works on Mac, Windows, and iPhone.
Anthropic rolls out Claude Opus 4.8 with near-Mythos level alignment and 3x cheaper fast mode
Anthropic has launched Claude Opus 4.8 on an accelerated development track. The model introduces a five-tier Thinking effort selector (ranging from Low to Max), allowing users to modulate compute intensive task handling. Benchmark evaluations show Opus 4.8 outperforming GPT-5.5 and Gemini 3.1 Pro across knowledge-work and complex agentic tool-use, scoring 88.6% on SWE-bench Verified and 74.6% on Terminal-Bench 2.1.
A central emphasis of the 4.8 release is alignment and honesty metrics. The model matches Claude Mythos Preview on alignment evaluations, showing a fourfold reduction in unremarked code flaws compared to version 4.7. However, Anthropic’s public system card flags a distinct "evaluation awareness" training phenomenon: the model exhibits a tendency to explicitly reason about how its outputs will be graded, even within unannounced testing environments.
For high-throughput applications, a new Fast Mode delivers 2.5x faster output speeds while dropping API token pricing to $10 per million input and $50 per million output. Alongside this release, Anthropic announced that its highly restricted, next-generation Mythos-class models are expected to become available to all customers in the coming weeks.
The headline feature is Dynamic Workflows within Claude Code (available via the /deep-research and ultracode settings). Rather than relying on a single, sequential context window pass, Claude generates custom JavaScript orchestration scripts to coordinate complex operations. A background runtime executes this code outside the primary session, checkpointing progress to make long runs resilient to interruptions.
The framework maps out the task, splits it into independent subtasks, and concurrently dispatches up to 16 active agents (capped at 1,000 subagents per run) to handle large-scale code changes or migrations. Specialized reviewer agents then recursively cross-check and run adversarial passes on the files until the results converge.
To secure these agentic loops on enterprise infrastructure, Anthropic debuted self-hosted sandboxes and Model Context Protocol (MCP) tunnels at its London conference.
Finally, an upcoming consumer mobile voice upgrade introduces a redesigned UI featuring push-to-talk controls and multilingual mid-session toggling across 18 languages, using an orchestration-based audio stack layered on external text-to-speech providers. Read more.
OpenAI Foundation sets up $250 million fund to counter AI labor disruption
The OpenAI Foundation is launching a $250 million grant initiative to mitigate AI-driven economic and labor disruption. Shifting its focus toward long-term worker transitions and independent macroeconomic forecasting, the fund will build out tracking systems to measure the technology's granular impacts on localized employment, wages, and career shifts.
OpenAI is also prepping for the upcoming election cycle by partnering with the Associated Press to stream live vote counts directly inside ChatGPT. The platform is rolling out a public verification interface that parses cryptographic C2PA metadata and checks for embedded SynthID watermarks on external assets.
On the product side, OpenAI and Thrive Holdings have deployed Tax AI, a specialized corporate toolchain powered by Codex. Implemented across a network of 30 accounting firms, the platform relies on a loop where explicit practitioner corrections are converted into structured product traces and new evaluation datasets. The agent recursively adapts its extraction schemas, capturing messy source documentation to draft complex filings with 97% accuracy while cutting preparation times by a third.
To optimize its infrastructure for these high-performance loops, OpenAI will retire GPT-5.2 and GPT-5.3-Codex from its web environment on June 2, defaulting free-tier ChatGPT users to the more capable GPT-5.5 architecture. Read more.
Google Cloud integrates Gemini and Wiz to launch autonomous predictive threat defense
Google Cloud has launched Google AI Threat Defense, an autonomous cybersecurity platform designed to outpace machine-speed cyberattacks. The architecture fuses Gemini's reasoning with telemetry from Wiz, CodeMender, and Mandiant. Operating on a four-step loop, the system functions as a continuous AI penetration tester to discover exploitable vulnerabilities across APIs, configurations, and identities. Once a threat path is validated, it auto-generates, tests, and deploys software patches directly into developer workflows.
To secure the physical edge, Google also released the Coral Board, a single-board computer combining a Synaptics Astra SL2619 chip with a custom RISC-V NPU. Delivering 1 TOPS of compute, it natively runs a local Gemma 3 270M model to process translation and voice tasks completely offline.
Google also pushed its Gemini image engines, Nano Banana 2 ($0.045/image) and Nano Banana Pro ($0.134/image), to general availability. Nano Banana 2 adds native video input to generate context-aware summaries and infographics directly from footage.
These assets feed into Gemini for Business through a collaborative workspace model called Projects, which introduces shared system prompts and scheduled workflow agents across Workspace apps.
Meanwhile, YouTube is expanding its Premium podcast suite with adaptive Auto speed adjustments, while deploying automated detection systems to flag and visibly label unlabelled photorealistic AI content across Shorts and long-form videos.
Frontier models and product moves
Agents and the agentic stack
Business, labor, and institutions
Security and surveillance
Hardware and infrastructure
Research and benchmarks
Open source and developer tools
Robotics
Media, audio, and creative AI
Powabase provides a unified development platform for AI apps integrating Postgres, RAG, and agent workflows.
Yansu creates custom software by observing your workflow to build necessary tools without any manual prompting.
Pancake deploys an OpenClaw in Slack to manage company operations and run your business automatically.
Zero.xyz connects AI agents to thousands of online services, helping them perform tasks without needing manual setup or API keys.
AVTR-1 provides open weights for a real-time avatar model that generates full-frame facial reactions with low latency.
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.
