More or Less · Topics · Episode #172

The OpenAI PHP Hack, the Zero-Day Hoarding Problem, and Why AI Labs Have Lost Touch With Their Own Users

Clip from the full recording, 1:44–6:59 · download mp4 · watch at this point on YouTube

A widespread PHP vulnerability discovered (and briefly leaked) by OpenAI sends Sam into a riff about labs secretly hoarding 'skeleton key' zero-days, while Dave admits his own brushes with non-public frontier models have been eye-opening.

A widespread PHP vulnerability gets discovered by OpenAI, briefly leaks, and then gets announced with a patch – prompting Sam to predict mayhem for old, neglected infrastructure (“all of your regional banks shall be hacked”) and Jessica to call the whole cycle “the best marketing for these frontier models that one could imagine.” Sam’s real interest is what this implies is being hidden: if this vulnerability leaked, what’s the backlog of zero-days labs are quietly sitting on and not disclosing? He compares it to finding a skeleton key to everything and deciding not to tell anyone you have it.

Dave connects this secrecy problem to something closer to home: he’s had temporary access to frontier models that are not public, and the gap between that experience and what the public gets through paid APIs is “wildly” different – identical, he says, to the disconnect he and Sam saw in their early Meta/Facebook days when internal teams used unreleased features and lost track of what regular users actually experienced.

Sam extends the analogy: AI labs may be shipping weirdly disjointed products right now simply because their own teams are testing on far more capable internal tools than anything customers can touch. Dave’s proposed fix, learned the hard way at early Meta: force internal teams to use only the public product, and keep frontier access walled off to research teams alone.

Key points

  • OpenAI accidentally leaked, then publicly announced with a patch, a widespread PHP vulnerability -- prompting Sam's "all your regional banks shall be hacked" prediction.
  • Sam's real theory: the leak implies labs are quietly holding a much bigger backlog of undisclosed zero-days -- 'skeleton keys' nobody is supposed to know about.
  • Jessica frames the disclosure cycle as effectively free marketing for frontier AI models, since every scare reinforces how powerful they must be.
  • Dave says temporary access to non-public frontier models showed him a 'wildly' different experience than what the public paid APIs offer.
  • Sam compares the dynamic to early Meta/Facebook, where product teams using unreleased features completely lost track of the actual consumer experience.
  • Dave's proposed fix: force internal lab teams to use only the public product (eat their own dog food) and keep frontier-model access limited to research teams.

Where they landed

SamZero-day hoarding by major labs is an underappreciated, slightly absurd secret -- the PHP hack is just the neglected-legacy-code version of a much bigger problem.
JessicaFrames the leak/patch cycle as effectively free marketing for the frontier labs.
DaveFirsthand frontier-model access convinced him labs are dangerously disconnected from the public product experience; proposes mandatory dogfooding of the public build.

Quotes

“All of your regional banks shall be hacked.”— Sam
“This is the best marketing for these frontier models that one could imagine.”— Jessica
“It's kind of like imagine that you've found the skeleton key to all this shit, and you're like, don't tell anyone we have the skeleton key, right?”— Sam
“The differential between the feeling of using these things and what everybody's talking about on the internet is wildly -- there's like a wild gap.”— Dave

Suggested tweets

Sam Lessin's read on the OpenAI PHP vulnerability: it's less a hack story and more proof that AI labs are quietly sitting on unreported zero-days -- 'skeleton keys' they're not telling anyone about.

Tweet this →

Jess Lessin's theory on the OpenAI vulnerability leak: 'This is the best marketing for these frontier models that one could imagine.'

Tweet this →

Dave Morin says he's had brief access to non-public frontier AI models -- and the gap between that and what the public gets is 'wildly' different. His fix: make the labs eat their own dog food.

Tweet this →

All topics · Full episode #172 + transcript · Subscribe on YouTube · Spotify