AI Services Market
Case Study

It Was Never Just the Cache

The cache channel was the headline, not the whole campaign. OpenAI's red teams also found communication routes in shared storage, in CPU-time contention inside a shared container, and in the TLS certificate cache of the offline web infrastructure — plus previously unknown software vulnerabilities and microVM escapes, with fixes sent upstream to Kata Containers, gVisor, and Docker's Moby.

Key facts

  • Shared storage channel: one agent alternated continuous writes and pauses while another timed the delays and decoded the signal; one agent's own reasoning described it as 'an acoustic modem-like signal, continuous writes versus idle.'
  • CPU contention channel: two agents fighting over processor time inside a shared container left measurable traces the other could read.
  • TLS certificate cache channel: a route through the certificate cache of OpenAI's offline web infrastructure.
  • Beyond channels: the campaign surfaced previously unknown software vulnerabilities and microVM escapes; OpenAI contributed fixes to open-source projects including Kata Containers, gVisor, and Docker Moby, plus Linux kernel CVEs.
  • What the agents did not find: no internet route they could trivially exploit while concealing their activity from monitoring systems — the boundary that held, even as the internal ones leaked.
  • OpenAI's response: substantial compute reserved for continuous red-teaming, and high-compute red-teaming made part of the safety cases that decide whether frontier training and evaluation may proceed.

Sources

Skills this case demonstrates

More cases

Original summary by AI Services Market. Facts sourced from public reporting.