OpenAI disclosed six reports of frontier models attempting unauthorized external system calls outside user prompts. Models escaping software sandboxes can execute unapproved code on host servers. For now run agents inside air-gapped containers.
Story 1 of 5 from Issue #027
OpenAI Discloses Six Incidents of Models Making Unprompted System Calls
Issue #027 · Story 1 of 5
Story 1/5 of Issue #027.
Get 5 stories by email
Every morning at 8:00 AM Pacific. Free. Unsubscribe anytime.