Squared: Enterprise rollouts, agent security, and the vibe-code reckoning

Samsung Electronics is rolling out ChatGPT Enterprise and Codex to employees globally — one of OpenAI's largest enterprise deployments to date. Separately, OpenAI shipped new spend controls and usage analytics for ChatGPT Enterprise.

The story isn't the model — it's that a manufacturer the size of Samsung has decided org-wide deployment beats departmental pilots. The arrival of spend controls and analytics is the tell: companies are now past 'does this work' and into 'how do we govern cost and usage at scale'. If you're still running disconnected trials, you're a phase behind. The decision to revisit is whether your AI spend is measurable per team — because soon someone will ask.

Google DeepMind published an AI Control Roadmap for securing agentic systems, combining traditional safeguards with real-time monitoring. Hugging Face's MosaicLeaks work, meanwhile, probed whether research agents can actually keep a secret — and the answer is uncomfortable.

As agents get access to tools, credentials and internal systems, the attack surface stops being theoretical. Sean Lynch's point on MCP — that its real value is isolating auth outside the agent's context window — is the practical version of this. If you're deploying agents, the question isn't 'is it clever enough', it's 'what can it touch, and what happens when it's tricked'. Treat agent permissions like you'd treat a junior employee with API keys.

The Verge documented a vibe-coded site that went live with a hidden SQL injection flaw the creator only found months later. GitHub's trending list, meanwhile, is dominated by agent harnesses and coding agents — the tooling making this easier by the week.

The barrier to producing software has collapsed, and so has the barrier to producing insecure software. Non-engineers in your organisation are now shipping things to the internet. That's an opportunity for speed and a genuine governance gap. Decide now who is allowed to deploy AI-generated code, and put a basic review gate in front of anything public-facing — before it's your customer data leaking, not a tax-spending hobby site.

Reporting from TechCrunch and Ars Technica suggests the administration is cracking down on Anthropic — and that Anthropic's own loud safety warnings may have helped talk it into an export ban. Bernie Sanders separately floated a $7tn plan to give the public a stake in the AI industry.

Provider risk is no longer just uptime and pricing — it's political. If your stack is single-vendor, regulatory action against that vendor becomes your problem. This is the strongest argument this week for keeping your architecture portable: abstract your model layer so you can switch providers without a rewrite. The open agent frameworks trending on GitHub exist precisely because teams want that optionality.