In partnership with
> SPOTLIGHT
WHAT MATTERS TODAY
OpenAI introduced new GPT-Rosalind capabilities for life sciences research, including the Life Sciences Research plugin, the Life Sciences NGS Analysis plugin for Codex, biological file viewers, expanded trusted access for qualified organizations globally, and Novo Nordisk as a partner.
GPT-Rosalind shows Codex being specialized into a domain workbench: model reasoning, research tools, biological files, evidence handling, provenance, and access controls inside one workspace. For expert workflows, the model is only one layer. The real value comes from packaging it with the tools, data, and permissions users need to work in a controlled way.
Anthropic analyzed 832 banned accounts tied to malicious cyber activity from March 2025 to March 2026, then mapped observed behavior to the MITRE ATT&CK framework. The report shows AI appearing in cyber operations that go beyond simple phishing or script generation, reaching deeper into multiple stages of attack workflows.
Agent risk is becoming workflow risk. When AI helps a user move through reconnaissance, tooling, execution, and follow-up, safety cannot be judged from a single prompt. It has to track the chain of work: what operation the user is trying to complete, what tools are being used, and whether intent changes across steps.
Meta introduced Meta Business Agent and Business Agent Platform, bringing agents into WhatsApp and Instagram for customer-facing workflows such as answering questions, recommending products, booking appointments, qualifying leads, handing off to humans, and helping close sales. The platform layer can connect to systems such as Shopify, Zendesk, and Shopee, with enterprise controls, guardrails, and measurement.
Meta’s planning to monetize through WhatsApp Business Premium tiers and token-based usage for large businesses. Meta is putting agents where customer intent already exists, then charging when those agents become real sales and support infrastructure.
> SIGNAL HEADLINES
CAPTURE THE SHIFT
Google introduces Gemma 4 12B for local multimodal AI: Google introduced Gemma 4 12B, a unified multimodal open model under an Apache 2.0 license. Google Gemma says it can run locally with 16GB of VRAM or unified memory, pushing capable AI down into laptops and edge workflows alongside the cloud frontier race.
OpenAI proposes a frontier safety blueprint: OpenAI is pointing policymakers to a frontier safety blueprint after the US cyber executive order. Frontier safety is moving toward cyber-risk evaluation, trusted early access, and government coordination.
AI could expand the amount of work: Aaron Levie argues that AI could create more software projects, more systems to maintain, and more sales and marketing throughput for companies to staff. This is not proof that AI will increase jobs everywhere; it reframes the operator question around what companies do when software becomes cheaper to create.
Anthropic builds out the services channel for Claude: Anthropic introduced a Services Track and Partner Hub after more than 40,000 firms applied to join the Claude Partner Network and more than 10,000 consultants earned Claude certification. Enterprise AI adoption is depending more heavily on implementation capacity.
GitHub makes VS Code more agent-first: GitHub says VS Code's Agents window is available in Stable preview, with remote agents over SSH or Dev Tunnels, continued Agent Host Protocol work, and sessions that can sync across machines. Coding agents are starting to look like durable work sessions rather than prompt boxes.
ChatGPT Enterprise Sites is stronger than early X reactions suggested: Simon Smith says his ChatGPT Enterprise account now has Sites with a relational database, file storage, environment variables, versioning, access controls, and server-side logic. This is a claim-by-source: ChatGPT business environments are moving closer to app hosting, data, permissions, and automation.
Lee Robinson pushes back on the “one builder role” narrative: Lee Robinson argues that AI and coding agents make code generation easier, but strong engineering, product, and design ownership still matter for managing complexity and avoiding low-quality shipping. AI may compress some role boundaries, but deep ownership still matters.
Microsoft AI says MAI-Thinking-1 is competitive on reasoning and coding: Microsoft AI says MAI-Thinking-1 is a 35B-active / 1T-total MoE reasoning model built from scratch, with claimed competitive performance on STEM reasoning and coding tasks. Microsoft is continuing to signal its own model-stack credibility.
Sundar Pichai says Alphabet raised $45B for the AI opportunity: Sundar Pichai says Alphabet's AI-linked equity offering was oversubscribed, raised about $45B, and has another $40B planned through an at-the-market program starting in Q3, including a $10B Berkshire Hathaway investment. AI infrastructure is becoming a capital-markets story.
Claude Code changes its dynamic workflows trigger to ultracode: ClaudeDevs says Claude Code changed the explicit trigger word from workflow to ultracode after feedback, while still allowing natural phrasing such as “use a workflow for this.” Agent products need more precise control language when a single word can trigger heavier orchestration.
> ONE PRACTICAL USE OF AI TODAY
Turn self-service analytics into a governed AI workflow
Anthropic says 95% of its business analytics queries are automated through Claude with roughly 95% aggregate accuracy. Analytics agents only work well when a company makes data legible: canonical datasets, metric definitions, source-of-truth routing, freshness checks, and validation loops.
In 30 minutes, pick one recurring analytics question in your business and turn it into a small governed workflow.
STEPS:
Pick one recurring question: revenue by channel, churn by segment, campaign ROI, support volume, sales pipeline, or content performance.
Name the canonical data source: the table, dashboard, sheet, warehouse view, or report that should win when sources conflict.
Write the metric definition clearly: what counts, what does not count, and which time window is valid.
Add a freshness rule: how recent the data needs to be before the answer is trustworthy.
Write the agent instruction: which source to use, which assumptions to avoid, which caveats to show, and when to refuse.
Add a validation check: compare the agent's answer against a known dashboard, a manual query, or a prior report.
Save the workflow as a reusable prompt, skill, SOP, or internal note.
EVALUATE:
If the agent gives a plausible answer but cannot cite the source, fix source routing first.
If the source is right but the number is wrong, tighten the metric definition.
If the number is right but the caveat is missing, add refusal and uncertainty rules.
If another teammate cannot reuse the workflow, it is still a chat answer, not an analytics system.
> PRESENTED BY BELAY
When Did Your Business Start Running You?
What started as ownership turned into obligation.
Now you’re in every meeting, decision, and channel… not because you want to be, but because things stall without you.
It’s not a capacity issue. It’s a structure issue.
The Freedom Framework shows you how to rebuild work flows, so you can step back without things breaking down.
BELAY U.S.-based Assistants help make that real by bringing ownership to execution, so your business doesn’t rely on you to function.
> WORTH READING
ANALYSIS & THESIS
An essay from The Free Press argues that AI coding tools can expand software ambition instead of simply killing coding jobs. Read alongside Aaron Levie and Lee Robinson, it adds a lens on what kinds of work become worth attempting when code becomes cheaper to create.
TechCrunch maps the current browser alternatives as AI changes how users search, browse, and act on the web. The browser is becoming an agent distribution surface, where agents can observe context and execute tasks.
Anthropic's full report provides abuse patterns, MITRE mapping, and evidence for how AI cyber risk operates through actions and workflows.
