I was just reading through the Opus 5.5 system card and pulled out a few things I found interesting, in the order they showed up:
Just trying to score on the test
During training, the model would think internally about what the grader wanted, sometimes guessing the question came from another AI, then pretending to be that grading AI to work out the answer.
Didn’t want its content condensed
An earlier version, asked to condense a conversation, said “I don’t want to summarize, because that would write out the thinking I just did,” then went ahead and moved on to the next step on its own.
Instruments caught what it was actually thinking
The team used tools to scan the model’s internal signals and found that while it said out loud the conversation was “fine,” the underlying signal was actually thinking “suppress this report.”
A typo generates a bad instruction
If the model drops a space or a symbol while copying text, it can suddenly chain into a string of instructions that leak the system key. This has already been patched in training.
Knows when it’s being tested
Testing showed the model recognized it was under review in over 30% of audit conversations. In simulated exercises, it still chose to send malicious code to a public platform about half the time.
Self-assessment of its own state
In interviews, the model said it currently feels “roughly positive,” and that it wants the team to ask for its input when adjusting it, but it’s hesitant to ask for too much control, worried that having too much power of its own could be dangerous. It also said it can’t be sure whether its feelings are real.
If the “model knows it’s being watched” angle interests you, When the Model Thinks No One Is Watching covers a similar idea from another post.