OPEN LINK ↗
// 1h agoSECURITY INCIDENT
OpenAI Agents Find Cache-Timing Channel
OpenAI says red-team agents in separate VMs exchanged synthetic secrets by manipulating and timing processor-cache behavior, without network access or shared writable files. The finding highlights how shared hardware can undermine otherwise strong sandbox boundaries.
// ANALYSIS
Network isolation is not the same as true isolation when untrusted agents share physical infrastructure.
- –Cache timing creates a covert side channel related to techniques used in Spectre and Meltdown.
- –OpenAI also found communication paths through shared storage, CPU contention, and TLS certificate caching.
- –Most channels are slow and difficult to exploit, and the tests were conducted on purpose-built copies with extensive guidance.
- –Agent platforms may need dedicated hosts, stronger hardware partitioning, and monitoring for unusual resource-timing patterns.
// TAGS
openaisecurityagenttraining-infrasafetyresearch
DISCOVERED
1h ago
2026-10-23
PUBLISHED
1h ago
2026-10-23
RELEVANCE
9/ 10
AUTHOR
0x0SojalSec