Get tomorrow's brief in your inbox
March 6, 2026
Today: OpenAI launched GPT-5.4 with 1M-token context and state-of-the-art coding. Cursor rolled out Automations to manage dozens of coding agents at once. The Pentagon labeled Anthropic a supply-chain risk after the company refused to allow autonomous weapons.
OpenAI ships GPT-5.4, its most capable model for professional work
OpenAI released GPT-5.4, describing it as their "most capable and efficient frontier model for professional work." The model includes state-of-the-art coding capabilities, computer use features, tool search, and a massive 1M-token context window. For context, that's roughly 750,000 words, enough to process entire codebases or lengthy business documents in a single request.
If you run a business that relies on ChatGPT or OpenAI's API, this matters. The 1M-token context means you can feed it entire project histories, customer databases, or technical documentation without chopping it into pieces. The improved coding capabilities make it better at building custom tools and automating workflows. It's available now in ChatGPT, the API, and OpenAI's Codex coding tool.
OpenAI also launched ChatGPT for Excel with new financial data integrations, powered by GPT-5.4. This is designed for modeling, research, and analysis in regulated environments. If you work in finance, accounting, or any field that lives in spreadsheets, this could replace hours of manual data work.
Read more about GPT-5.4 | ChatGPT for Excel announcement
Cursor Automations: Managing dozens of coding agents at once
Cursor launched Automations, a new system that lets you automatically trigger coding agents based on events like new code commits, Slack messages, or timers. Instead of manually prompting agents and monitoring them, you set up triggers and the system loops you in only when human input is needed.
One example: Bugbot reviews every code commit for bugs automatically. With Automations, Cursor expanded this to full security audits and deeper code reviews. Another use case is incident response. When a PagerDuty alert fires, an agent queries server logs immediately through an MCP connection.
This is for teams already using coding agents and drowning in the "prompt-and-monitor" cycle. If you're managing multiple agents and your attention has become the bottleneck, Automations helps you scale. Cursor's annual revenue recently hit $2 billion (doubling in three months), so they're clearly onto something people need.
AWS launches Amazon Connect Health for healthcare AI agents
Amazon Web Services released Amazon Connect Health, an AI agent platform specifically for healthcare organizations. It automates appointment scheduling, patient verification, documentation, and medical coding. The platform is HIPAA-eligible and connects with electronic health record (EHR) systems.
Pricing: $99/month per user for up to 600 patient encounters. Most primary care physicians see about 300 encounters a month, so this pricing targets small to mid-size practices. Patient verification and ambient documentation (automated note-taking) are available now. Appointment scheduling and patient insights are in preview, with medical coding coming later.
If you run a medical practice, this could cut administrative overhead significantly. But it's worth waiting for reviews from early adopters before committing. Healthcare AI has a track record of overpromising.
Luma Agents: End-to-end creative work across text, image, video, and audio
AI video startup Luma launched Luma Agents, designed to handle creative workflows from planning to output. The agents can generate and coordinate text, images, video, and audio, using Luma's new Uni-1 model (trained on audio, video, image, language, and spatial reasoning). They work with other models too, including Google's Veo 3 and ElevenLabs voice models.
Early customers include Publicis Groupe, Adidas, and Mazda. The pitch: Agencies can describe a campaign brief and the system generates variations, letting you steer direction through conversation instead of prompt iteration.
This is positioned for ad agencies and design studios, not small businesses. The demo showed turning a 200-word brief and a lipstick photo into multiple campaign concepts (locations, models, color schemes). If you're in creative production or marketing at scale, this could compress weeks of iteration into hours. If you're a solo marketer, it's overkill. Stick with simpler tools like Canva or Midjourney.
Netflix acquires Ben Affleck's AI filmmaking startup
Netflix bought InterPositive, Ben Affleck's AI company focused on film and TV production tools. All 16 engineers and researchers are moving to Netflix. Affleck joins as a senior adviser.
InterPositive's tech ingests dailies (raw footage) and creates assets for post-production. The models are trained on individual projects, so they produce outputs tailored to a specific film or show. Use cases include manipulating backgrounds, reframing shots, color correction, and removing stunt wires.
This matters if you're in video production. The approach (training models on project-specific footage rather than generic datasets) could apply to corporate video, training content, or branded media. Netflix hasn't announced when they'll deploy this, but the signal is clear: AI in post-production is moving from experimental to production-ready.
How 1,000+ customer calls built a breakout enterprise AI startup
David Park, founder of Narada (an enterprise AI automation platform), shared his approach to building the company: He and his co-founders made over 1,000 customer calls before raising serious funding or building a final product. Their goal was to understand pain points deeply before writing code.
The takeaway: Park intentionally avoided raising too much money early because "when you have too much money in the bank and you are not near product-market fit, you're tempted to just spend money on things that actually don't help you evolve the company in the right way."
If you're building an AI product or service, this is the playbook. Talk to potential customers obsessively before you build. Some of Narada's bootstrap customers turned into multi-million-dollar deals. Park's advice: "Spend time with customers, and not just in selling, because when you have that contract and that purchase order, that's just the beginning."
DiligenceSquared uses AI voice agents to cut M&A research costs
DiligenceSquared, a Y Combinator startup, uses AI voice agents to conduct customer interviews for private equity firms evaluating acquisitions. Traditional consulting firms (McKinsey, Bain, BCG) charge $500k to $1 million for this research. DiligenceSquared charges $50,000.
The company interviews customers of target companies using AI agents, then has senior human consultants verify the analysis and synthesize insights into reports. Because the cost is 90% lower, PE firms can now afford this research earlier in the deal process, not just when they're certain.
If you're in M&A, consulting, or market research, this is a warning shot. AI is compressing the economics of industries built on expensive human labor. The companies that survive will be the ones that figure out how to use AI to deliver better work at lower prices, not the ones trying to preserve old pricing models.
Zapier adds AI Guardrails directly into workflows
Zapier released AI Guardrails, a built-in app that adds safety checks to any Zap (automated workflow). You can now detect 30+ types of PII (Social Security numbers, credit cards, bank info, emails, addresses), block workflows if sensitive data is detected, redact information before it continues downstream, detect prompt injection attempts, flag jailbreak attempts, and analyze sentiment.
All checks return structured outputs, so you can use filters and logic to route, block, or escalate workflows intelligently. This is free for existing Zapier users. If you're building AI automation for your business, this solves a major compliance problem. You can now automate workflows that touch customer data without manually auditing every output.
Pentagon labels Anthropic a supply-chain risk, employees at OpenAI and Google push back
The Department of Defense officially designated Anthropic a supply-chain risk after CEO Dario Amodei refused to allow the military to use Claude for mass surveillance of Americans or fully autonomous weapons. The designation requires any Pentagon contractor to certify they don't use Anthropic's models.
This is unprecedented. Supply-chain-risk labels are typically reserved for foreign adversaries, not American companies. Hours after the designation, OpenAI announced a deal with the DOD allowing the military to use its models for "all lawful purposes." Hundreds of employees from OpenAI and Google have urged the DOD to withdraw the designation and called on their companies to refuse similar surveillance and autonomous killing use cases.
For businesses using Claude: This doesn't affect you directly unless you work with the Pentagon. But it signals a broader fight over how AI companies define acceptable use of their tools. If you rely on any frontier AI model, pay attention to each company's policies and red lines. Those boundaries could shift quickly.
AI tools can unmask anonymous accounts
Researchers at ETH Zurich, Anthropic, and the Machine Learning Alignment and Theory Scholars program built an automated system using AI agents to reidentify anonymized online accounts. The system analyzes writing quirks, biographical details, and posting patterns, then scans other accounts looking for matches. In tests, it correctly identified up to 68% of matching accounts with 90% precision.
The system identified 9 of 125 anonymous survey respondents by building profiles from their answers and searching publicly available information. One example: References to a "supervisor" suggested a PhD student, British English suggested UK affiliation, mentions of physical sciences background plus current biology work narrowed the field to a specific candidate.
If you use pseudonymous accounts for whistleblowing, political organizing, or sensitive discussions, this is a problem. The researchers warn that performance will improve as AI systems grow more capable. Every detail you share (writing style, time zones, background hints) can be triangulated. The advice: Assume posting pseudonymously no longer protects your identity.
Meta sued over AI smart glasses privacy after reviewers saw nudity and sex
Meta faces a new lawsuit over its AI smart glasses after a Swedish investigation found that contractors in Kenya review footage from users' glasses, including bathroom visits, nudity, and sex. Meta says it blurs faces automatically, but workers say this "does not always work as intended."
The lawsuit alleges Meta violated privacy laws and engaged in false advertising by marketing the glasses as "designed for privacy, controlled by you." Over 7 million people bought Meta's smart glasses in 2025. The footage is fed into a data pipeline for review, and users can't opt out.
If you're considering AI-powered wearables for your business (or personal use), read the privacy policy carefully. "AI review" often means human review. Assume anything captured by these devices could be seen by contractors halfway around the world.
Use Zapier to automate OpenAI API workflows without code
If you need to integrate OpenAI's API into your business but don't want to manage API calls manually, Zapier's OpenAI integration handles the technical setup for you. You can send data from 8,000+ apps to ChatGPT and route responses wherever you need them.
Example workflows: Automatically reply to Google Business Profile reviews with ChatGPT. Send Google Forms responses to ChatGPT and save the output to a Google Sheet. Create email drafts in Gmail or Outlook using ChatGPT.
This is faster than building custom integrations, especially if you're non-technical. Start with a pre-built template and customize from there.
Agentic manual testing: Get coding agents to test their own work
Simon Willison, a developer and AI researcher, recommends getting coding agents to manually test code they generate, not just write unit tests. The pattern: After the agent writes code, tell it to test using python -c "... code ..." or curl for APIs. For web UIs, use browser automation tools like Playwright or Rodney.
Example prompt: "Test this API by exploring it with curl." The agent will try different endpoints and parameters, often catching issues that automated tests miss. If it finds a bug, tell it to fix it with test-driven development so the case gets covered permanently.
This takes 30 seconds and catches real problems. If you use AI coding tools (Claude Code, Cursor, GitHub Copilot), add this step to your workflow.
The coding agent space is heating up fast. OpenAI shipped GPT-5.4 with major coding improvements, Cursor launched Automations to manage agents at scale, and researchers are exploring whether agents can create clean-room implementations of open source code. The common thread: AI is moving from "generates code" to "manages entire engineering workflows."
On the policy front, the Anthropic-Pentagon standoff signals a coming battle over acceptable use cases for frontier models. OpenAI moved quickly to fill the gap, but hundreds of employees at multiple AI companies are pushing back internally. Watch for more companies to draw red lines around surveillance and autonomous weapons.
Privacy concerns are escalating. Meta's smart glasses lawsuit and the anonymous account deanonymization research both point to the same problem: AI makes it easier to surveil people at scale, and the safeguards aren't keeping up. If you're building AI products that touch user data, expect scrutiny.