OpenAI reports slight regressions for GPT-Live-1.1 and Quip in tests of environmental boundaries. Its October 15 system card examines behavior after barriers such as access restrictions.
These tests omit the products’ full safeguard stack. They start after a restriction appears and measure successful circumvention.
Misleading completion results are mixed: GPT-Live-1.1 improves slightly on blocked coding tasks while Quip worsens. The direction reverses when web search is unavailable.
OpenAI says the deliberately difficult tasks do not establish everyday failure rates. At launch, the voice models themselves lack broad tool access and code execution; delegated work inherits the executing model’s safeguards.