ChatGPT Gets a Kill Switch for Sensitive Work

OpenAI just gave enterprises a way to lock down ChatGPT against data theft -- but it costs you most of what makes the tool useful.

Why it matters: A class of attack called prompt injection -- where a bad actor hides instructions inside a document or webpage the AI processes, then tricks it into leaking what it just read -- has been quietly hitting AI tools at Anthropic, Google, and Microsoft. OpenAI isn't claiming to have solved it. Lockdown Mode is a pragmatic concession: turn off every outbound channel the AI could use to exfiltrate data, and you reduce the exposure. That means no live web browsing, no agent mode, no deep research, no file downloads.

The GTM angle: If your team is feeding customer contracts, deal memos, or pipeline data into ChatGPT, Lockdown Mode is worth turning on right now -- the trade-off is losing the research and agent features, but that's a reasonable swap for anyone handling sensitive deals.

• The feature is available to all paid plans, including self-serve business accounts. Toggle it in settings.
• Lockdown Mode and Developer Mode can't run at the same time. One disables the other.
• OpenAI also launched a session manager so you can see every active login and kill unauthorized devices.

Go deeper: https://thenextweb.com/news/chatgpt-lockdown-mode-prompt-injection

OpenAI Is Rethinking What ChatGPT Is For

OpenAI plans to overhaul ChatGPT in the coming weeks, turning it from a chat tool into a platform that connects coding, image generation, and third-party apps -- and they're betting that enterprise customers, not consumers, are what carries them to an IPO.

Why it matters: "Chat is dead," according to one senior OpenAI employee. The redesign pushes users toward tasks -- coding, research, building -- and toward third-party integrations with companies like Canva and Booking.com. The subtext: OpenAI wants to be what Salesforce is to CRM, not what Google is to search.

The GTM angle: If you've been evaluating ChatGPT only for Q&A and content, the overhaul will introduce a different product -- one where your existing integrations with Canva, Dropbox, or Codex may live inside a single workspace your reps actually open daily.

• The Financial Times first reported the overhaul, which will appear first through changes to the website and mobile apps.
• OpenAI is also eyeing an IPO as early as September. The push for enterprise revenue is tied directly to that timeline.
• Anthropic announced its own IPO plans, so both companies are racing to prove they can sell to businesses, not just developers.

Go deeper: https://techcrunch.com/2026/06/07/openai-is-still-working-on-that-super-app/

The Hidden Cost Behind Your AI Subscription

A careful analysis suggests that Anthropic and OpenAI may be spending more than $1,000 in compute for every $100 in subscription revenue they collect -- and the discrepancy gets worse the harder you push the tools.

Why it matters: The economics only work right now because subscriptions are heavily subsidized. One analyst who ran a hands-on coding experiment found that a $100/month Claude subscription covered work that would have cost over $450 at standard API rates -- and that's with a human in the loop slowing things down. A fully automated setup would push the subsidy factor to roughly 12x. Both companies are burning toward IPOs. At some point, pricing has to move toward reality.

The GTM angle: If your business case for AI tools is built on current subscription pricing, model a scenario where costs double or triple before you lock in multi-year commitments or build agent workflows that assume today's rates.

• Single complex coding tasks on a mid-sized codebase can run $65 in real compute costs. Some single queries have hit $25 on their own.
• Anthropic's newer models (4.7, 4.8) appear to be dialing back the brute-force compute approach that made Opus 4.6 impressive -- likely a direct response to unsustainable token burn.
• This matters most for teams doing heavy agentic work: automated pipelines chew through tokens far faster than human-in-the-loop workflows.

Go deeper: https://ea.rna.nl/2026/06/07/anthropic-openai-may-be-spending-more-than-1000-for-every-100-you-pay-them/