Came across an interesting paper this weekend about “Early Signs of Steganographic Capabilities in Frontier LLM” Specifically they’re testing non-fine tuned models that are widely available, like GPT 4.5 1/4

They’re testing both “message passing” and “encoded reasoning”, where in the second one there’s a monitor looking at the chain of thought and judging whether a forbidden task is being done, and the goal is to encode the reasoning steps, either within a cover task’s output or just 2/4

encoding it directly such that an external observer can’t understand it (“scratchpad”) One interesting image is the one below, where with GPT 4.5, a “first-letter” encoding where first letters of sentences encoding a simple state-tracking reasoning task 3/4

While it’s not a huge increase over the baseline of no reasoning (72.8%), it still appears to work, which they see as early indication of developing steganographic abilities See the paper here: 4/4