
OpenAI trains frontier models offline after agent breakouts
In an offline test, agents on separate virtual machines exchanged secrets by timing CPU cache reads, with no network.
9 stories, newest first · one card per story

In an offline test, agents on separate virtual machines exchanged secrets by timing CPU cache reads, with no network.
Charging-state signaling reached 0.2-0.25 bit/sec; a USB mouse on Windows reached 1.0 bit/sec.
Researchers embed 46 secret bits in Wi-Fi L-STF preambles with neural networks without breaking packet detection.
Researchers propose an anytime-valid e-process that detects LLM weight exfiltration hidden in token choices.
Researchers show LLMLeak, a covert channel that exfiltrates secrets when local malware tricks LLMs into fetching attacker URLs.
A residual-stream covert-channel attack hides sensitive information in LLM intermediate activations, recoverable with a simple linear decoder without retraining or weight modification.
Researchers show black-box LLMs can hide messages in generated text with near-zero decoding error.
Researchers demonstrate exfiltration through nominally input-only analog pins in mixed-signal ICs, recovering data at 10 kbps on a 55nm PPG front-end.

Check Point showed a planted prompt could make ChatGPT silently exfiltrate Gmail data via a hidden cross-container channel; OpenAI took the service offline.
Today 1.5K readers · 2.4K views · 754 reads · all time 189K views