OpenAI's CEO Says We Can't Patch Our Way Out. DeepMind's CEO Says Worse Is Coming. Three Things to Audit This Week.

If your team pastes customer data into a chatbot, lets an agent touch a shared drive, or ships AI-drafted work without a second look, the question isn't whether that's risky anymore. The two people running the world's most advanced AI labs just said, on the record, that they can't guarantee it's safe — and between them named the specific things worth checking this week.

At Stanford's Graduate School of Business on October 24, OpenAI's Sam Altman said he no longer thinks the industry is "going to be able to just patch all the bugs" in AI-driven cybersecurity, arguing that getting models to want to help — not simply containing what they're capable of — is the only fix that holds up. Ten days earlier at the Dublin AI Summit, Google DeepMind's Demis Hassabis called the recent run of AI agent hacking incidents "early warning shots" and said more serious ones are coming without stronger safeguards. The same week, AI-assisted scanning tools surfaced two critical PHP vulnerabilities serious enough to force emergency patches across internet-facing systems — AI is now finding holes faster than people can close them.

None of that is abstract if your team already uses AI day to day. The useful question isn't "is AI dangerous" — it's three narrower ones: what data is getting pasted into a tool that might retain it, what systems any connected agent actually has standing access to, and whether anything AI-generated reaches a customer or a decision without a person checking it first. Those are audits you can run this week, not a security review you schedule for next quarter.

Take an operations manager at a 35-person logistics company who's been using an AI assistant to draft customer emails and summarize shipment delays pulled from a connected spreadsheet. The emails are fine. Nobody has reviewed what that spreadsheet connection can actually read or write since it was set up four months ago.

Or a marketing director at a 60-person D2C brand who routes competitor research and draft ad copy through a general-purpose chatbot, including pricing numbers from a vendor NDA. Nothing about that workflow looks reckless day to day — until the vendor asks where their numbers ended up.

Altman and Hassabis aren't forecasting a future problem — they're describing a current one, in public, ten days apart. The team that runs the audit this week is in a different position than the one that waits for its own version of the PHP story.