SuperCareer Daily AI Brief: Wednesday, 15 July 2026

The AI news that moves your career — in 60 seconds a day.

☕ The 60-second version

OpenAI pushed three pieces today reframing enterprise AI around "useful work per dollar" — a concrete framework for measuring agentic AI ROI, plus real playbooks for how sales and data-science teams use ChatGPT Work and Codex.
GitHub shipped BYOK expansion for Copilot in JetBrains and a new MCP trust layer in Visual Studio, signaling that model-choice and MCP-server vetting are becoming standard IT-governance line items, not power-user tweaks.
Security researchers disclosed a Cursor 0-day and a Tailscale SSH root-access bug the same week Codex started encrypting sub-agent prompts — coding-agent security is now a live, not theoretical, professional skill gap.

🔥 Today's big story

OpenAI tells enterprises to stop counting seats and start counting "useful work per dollar"

OpenAI's new framework asks companies to measure agentic AI spend by output produced per dollar, not by licenses purchased — a direct challenge to how most orgs currently justify AI budgets.
The companion playbooks show ChatGPT Work and Codex already doing real deliverables: sales teams generating pipeline briefs, meeting-prep packets, forecast reviews, account plans, and stalled-deal diagnoses; data science teams producing root-cause briefs, impact readouts, KPI memos, and dashboard specs.
This is a template shift: the deliverable itself (the memo, the brief, the diagnosis) becomes the unit AI is judged on — which means professionals who can produce and vet those artifacts fast become the ones AI amplifies, not replaces.

👔 If your team can't yet point to a specific recurring deliverable (a weekly forecast memo, a KPI readout, a deal-stall diagnosis) that AI drafts for you, you're behind the ROI curve your leadership will start measuring against. Learn to frame your own output in "useful work per dollar" terms before your manager does it for you in a headcount review.

How to manage AI investments in the agentic era · How sales teams use ChatGPT Work · How data science teams use ChatGPT Work

Level up your career with SuperCareer. Daily 10-minute challenges, AI tutoring, and real workplace skills. Try today's challenge free →

📰 Also today

GitHub Copilot adds a trust layer for MCP servers and expands BYOK to all JetBrains tiers

Visual Studio's June update adds a trust layer specifically for vetting MCP servers before they touch your codebase — the first mainstream "is this AI tool safe" gate baked into an IDE.
JetBrains Copilot now lets every tier bring their own model key, decoupling "which AI you use" from "which subscription tier you're on."

👔 Engineering managers: MCP-server vetting is about to be an IT policy question, not a dev preference — get ahead of it before a rogue MCP server becomes a security incident on your watch.

GitHub Copilot in Visual Studio — June update · GitHub Copilot for JetBrains expands BYOK capabilities

Bonsai 27B: a phone-class model lands as coding-agent security cracks show

Bonsai 27B claims full on-device operation at 27B parameters — pushing capable AI further off the cloud and onto phones, which changes what "AI-enabled" means for privacy-sensitive roles.
The same week, a Cursor 0-day disclosure and a Tailscale SSH bug (TS-2026-009, permitting root access) surfaced, and OpenAI's Codex started encrypting sub-agent prompts — a coordinated signal that AI coding tools are now a real attack surface.

👔 If you use AI coding tools daily, treat prompt/config hygiene and MCP-server sourcing as part of your job, not IT's — a leaked sub-agent prompt or a compromised extension is now a resume-relevant incident, not an edge case.

Bonsai 27B: A 27B-Class Model that runs on a phone · Cursor 0day: When Full Disclosure Becomes the Only Protection Left · TS-2026-009: Insecure argument handling in Tailscale SSH · Codex starts encrypting sub-agent prompts

🛠️ Use this today — Write your own "useful work per dollar" memo

Pick one recurring task you or your team does weekly (a status report, a deal review, a KPI readout). Prompt: "Draft a [weekly forecast review / KPI memo / stalled-deal diagnosis] template for [your role/team] that an AI assistant could fill in from raw data each week — include the 5 fields a manager actually reads first." Use the output to show, in one page, exactly what AI now produces for you and how fast — that's the artifact your next budget or promotion conversation will run on.

⚡ The feed

Agents

OpenAI Codex sub-agents now encrypt their prompts by default, closing a leakage path between orchestrator and sub-agent.

Business

Tools

Research

Other

Google Images turns 25, marking a quarter-century of visual search as AI reshapes how people search.

📈 Skill of the day

Stop saying "I used AI to help." Start saying what it produced and how fast — a one-line "useful work per dollar" framing (task → time saved → deliverable) is becoming the actual currency in performance and budget conversations.

❓ FAQ

What does "useful work per dollar" mean in OpenAI's new framework?

It's a way to measure agentic AI ROI by counting actual completed deliverables (memos, briefs, diagnoses) produced per dollar spent, instead of counting licenses or seats purchased. OpenAI is pushing enterprises to adopt this as the standard for justifying and scaling AI investment in the agentic era.

What is ChatGPT Work and how are sales and data teams using it?

ChatGPT Work is OpenAI's workplace-focused product tier. Sales teams use it to generate pipeline briefs, meeting-prep packets, forecast reviews, account plans, and stalled-deal diagnoses; data science teams use it for root-cause briefs, impact readouts, KPI memos, and dashboard specs — all drafted from real work inputs.

Why does it matter that Codex now encrypts sub-agent prompts?

Multi-agent coding systems pass prompts between an orchestrator and sub-agents, and unencrypted prompts could leak sensitive instructions or data in that hand-off. Encrypting them closes a real attack surface as coding agents handle more production code and credentials.

What is Bonsai 27B and why does an on-device model matter for careers?

Bonsai 27B is a 27-billion-parameter model designed to run fully on a phone rather than in the cloud. For privacy-sensitive roles (healthcare, legal, finance), on-device AI means capable assistance without sending data off the device — a growing requirement in regulated jobs.

Join the SuperCareer AI career newsletter for your personalized roadmap.