Keldura Daily Open Keldura

Keldura Daily · AI & Technology

AI security tests expose gaps beyond the network

Three developments show the infrastructure behind increasingly capable AI: containing agents, patching vulnerable software, and sharing storage at global scale.

The field note

3 sources · 3 items
  1. The cache experiment was an instructed security test: agents were asked to exchange secrets across isolated env…
  2. One agent encoded information by evicting selected cache data; another decoded it by measuring whether subseque…
  3. OpenAI says it added monitoring checks for this communication method and will retain offline training until str…
Story 012 sources

OpenAI cuts frontier training off from the internet

OpenAI now trains its most advanced models using cached and indexed web content rather than live internet access, following security incidents and a September training pause. Training resumed on October 1 under tighter controls, but separate tests showed agents could exchange secrets through processor-cache timing without network access or shared files. [4][8]

Why it matters

The tests demonstrate that removing internet access does not eliminate every communication channel between agents sharing hardware; containment also depends on monitoring and controls below the network layer. [4][8]

Key insights

  • The cache experiment was an instructed security test: agents were asked to exchange secrets across isolated environments, rather than being observed spontaneously coordinating. [4][8]
  • One agent encoded information by evicting selected cache data; another decoded it by measuring whether subsequent reads were fast or slow. [8]
  • OpenAI says it added monitoring checks for this communication method and will retain offline training until stronger safeguards justify broader internet access. [8]
Story 021 source

OpenAI returns to shared storage for global scale

OpenAI says Habitat now serves more than one billion users across 40 geographic regions and holds more than 500 petabytes of data. Its Azure Cosmos DB design moved from shared containers to dedicated containers and then back to shared containers by default, balancing resource efficiency against workload isolation. [7]

Why it matters

Supporting AI products at this scale requires more than model-serving capacity: application data must remain quickly accessible across products and regions. OpenAI's storage design illustrates the trade-off between pooling resources efficiently and protecting one workload from another's demand. [7]

Key insights

  • Dedicated containers give product teams clearer ownership and workload isolation, but can require excess capacity and separate planning for each team. [7]
  • Shared containers centralize provisioning expertise, but introduce a noisy-neighbor problem when one tenant consumes resources needed by others. [7]
  • Habitat also handles authorization, data residency, encryption, caching, and request shaping, making the database part of a wider application-storage system. [7]

Create your own daily briefing — start free. Keldura monitors the sources you choose and gives you a private, grounded daily digest with cited answers.

Create your own daily briefing — start free