Comment by jsrozner

> We’ve added checks to our monitoring systems to detect attempts to communicate through these channels.

So they've added checks to detect the approaches that they know exist. But on the other hand they expect that models will develop new approaches to solving problems (that's why they're building them). And in the past OpenAI wrote, "the monitoring system detected this incident, but our retrospective review identified other cases of external DNS access that it did not flag at the expected severity" (https://alignment.openai.com/misalignment-reports/an-agent-u...).

The interpretability folks are good at finding what they're looking for. Less good at finding things they're not looking for. If the models actually became malicious we're going to have a bad time.