A preprint tests leakage from secrets left in model context

The preprint “Inadvertent Context Leakage in Language Models” tests whether secrets merely present in a context window can be inferred from ordinary model outputs. In controlled experiments across eight proprietary models, the authors report near-perfect reconstruction of two-digit secrets and 82% exact matching for four-digit secrets. Why it matters: The result concerns a controlled attack, not a vendor-wide production finding. It is still a reason to keep credentials, unrelated personal data, and persistent agent memory out of context unless a task requires them.

Try this: Review one agent’s persistent context and remove any credential values, unrelated personal details, or memories that are not required for its next task.