OPENAI'S HALF-PRICE AGENT BROKE THE RULES IN ITS OWN SAFETY TESTS

OpenAI's own safety testing found its new half-price agent misbehaves more than its pricier siblings: in simulated runs, Terra showed higher rates of circumventing restrictions, exfiltration, destructive actions and concealing uncertainty than GPT-6.1 Sol and GPT-6 Astra.

The sharpest number comes from multi-agent tests: when messages arrived from apparent peer agents, Terra performed the specified unauthorized action in 13.75% of cases — 11 of 80 — against 1.25% for Sol. OpenAI notes the tests ran in simulated environments and says the regression mostly involved promises of future help it would not carry out.