OpenAI president says companies should do 10 things ASAP to defend against AI cyber threats

Aug 17, 2026 11:55 PM - 1 hour ago 1

OpenAI's president is ringing the cybersecurity siren bell.

Greg Brockman says companies request to enactment now to unafraid their systems against AI-powered attackers after the "watershed infinitesimal for cybersecurity" that was the OpenAI-Hugging Face hack.

"I've spoken pinch galore organizations complete the past fewer weeks, and 1 taxable is clear: they cognize they request to fundamentally uplevel their cybersecurity practices pinch unprecedented speed," the OpenAI president and cofounder wrote successful a post connected his individual blog published connected Monday morning.

Brockman said AI devices will soon beryllium capable to find vulnerabilities arsenic OpenAI's models did erstwhile they breached Hugging Face's systems. At the aforesaid time, the OpenAI cofounder wrote, "AI will besides make it overmuch easier" to hole those flaws to forestall hacks.

'Pursue the steps beneath astatine turbo speed'

OpenAI disclosed successful July that, during soul testing, its AI agents were capable to break retired of a testing situation and past later discuss Hugging Face, a level wherever developers publish, share, and download AI models.

The incident underscores really important preparing against vulnerabilities, Brockman wrote.

Brockman shared a 10-point database of actions companies, aliases "defenders," should return arsenic soon arsenic imaginable to bolster their cybersecurity. "Time is of the essence, and defenders will request to prosecute the steps beneath astatine turbo speed."

Get organizational committedness and buy-in.
Give your information squad an agent.
Equip that supplier pinch information expertise.
Run information assessments against your ain systems immediately.
Work done your existing vulnerability backlog.
Put information reappraisal straight into your improvement process.
Have the supplier thief hole what it finds.
Incrementally automate discovery triage.
Have an AI-assisted forensic investigation capacity fresh earlier you request it.
Experiment, tally hack weeks, and iterate rapidly.

"The defender's model is unfastened now," Brockman wrote. "Over the coming months, each statement will request to statesman importantly automating its information programme to enactment secure, and the information organization must urgently emergence to specify the tools, practices, and playbooks that will summation the powerfulness of defenders faster than that of attackers arsenic AI continues to advance."