← @bgSponge

On July 25, we hacked OpenAI.

Summary

s1r1us shows how a libheif heap overflow and an OpenAI SSO flaw allowed HacktronAI to take over employee ChatGPT accounts and access connected services like GitHub and Slack in under 72 hours.

  1. libheif vulnerabilities enable RCE.

    A heap overflow in libheif, which was fixed upstream but not flagged as security-relevant, allowed RCE on the OpenAI community forum.

    The image package libheif had a known vulnerability

  2. SSO flaws expand attack surfaces.

    An OpenAI SSO vulnerability turned forum access into ChatGPT and Codex account takeovers, potentially exposing connected Outlook, Gmail, and Slack data.

    This made the potential impact much larger than ChatGPT alone

  3. AI agents accelerate exploit development.

    Opus 4.8 identified the libheif vulnerability and built a partial exploit, which Opus 5 later adapted to achieve RCE on a test instance.

    AI is reducing the amount of scarce expertise needed to develop exploits

  4. Obscure libraries threaten major platforms.

    Harsh Jaiswal notes the libheif issue is an xkcd #234 scenario, affecting Slack, Meta, and GitHub Ent through a single obscure image library.

    one obscure image library beneath a huge number of apps

Read the original ↗Save to your Sponge

Summarized with Sponge. Quotes are the author’s own words.

More from @bg

All readings →