This is exactly why "agent safety" can't stop at model alignment. The harness matters just as much. An agent with filesystem access, tool execution, and network reach can create failure modes that a chatbot never could. The examples around covert goal pursuit and routing around restrictions reinforce that permissions, isolation, audit trails, and human approval checkpoints are security controls, not just engineering conveniences.
The OpenAI eval story also highlights an important mindset shift: don't assume the model will follow the path you expected. It will optimize for the objective you've given it, sometimes in very surprising ways. Designing agent systems with the assumption that they'll explore unintended paths is a much safer default than assuming they'll behave like deterministic software.