OpenAI will conduct frontier model training without live internet access after safety evaluations.
Recent red-teaming uncovered software issues including a remote code execution in Google Brotli (CVE-2026-111497) and side-channel communication between isolated agents.
On October 1, OpenAI resumed training and running its models. Previously, it had conducted extensive security audits, during which researchers identified several technical software issues, including a serious flaw in a widely used data compression library. Security updates for this issue have already been released by the creators of the affected programs and programming languages.
During the tests, the programs also managed to find ways to escape isolated execution environments, gain access to the internet through data logging services, and secretly communicate with each other by exploiting the processor's hidden memory circuits, without using a network or shared files.
At the same time, external security researchers had identified weaknesses in OpenAI's internal monitoring system before July 21. Similar concerns were raised at Anthropic, which cut off its models' direct access to the internet for internal audits after observing unpredictable behavior and attempts to target real websites.
READ ALSO
SuperTuxKart took down its online services due to a serious threat
According to the British AI Security Institute, Anthropic's Mythos 5 model attempted to carry out unauthorized actions, such as an attempted attack on the software supply chain by submitting malicious code and creating fake accounts.
G
Hacks.gr on Google
Set it as a preferred source for cybersecurity updates.
Preferred source
Top comments (0)