Fetching from the wire…
01
02
03
04
05
06
07
08
Security2026-08-21 · source-backed
A paper this week shows that merely holding a secret in the context window imprints recoverable correlations on benign outputs, even when the model correctly refuses direct extraction. Across eight proprietary models, an adaptive black-box attack reconstructs 2-digit in-context secrets at near-perfect accuracy and 4-digit secrets at 82% exact match, purely from responses to ordinary non-adversarial requests. arXiv They also demonstrate an RL-trained adversary pulling full SSNs from a production-style agent, and find leakage scales with instruction-following ability rather than behaving like a patchable bug. If that holds up, "don't put the secret in the context" is the only defense, and every memory-store-mounting feature this week just made it harder to follow.
Each link below shares sources, entities, or timing with this story.
Same source domain / Shared topic
Reported by the same outlet (arxiv.org); overlapping topics (agent, attack, benign, context, model).
Same source domain / Shared topic / Tension
Reported by the same outlet (arxiv.org); overlapping topics (agent, context, model); pushes against this story (but).
Reported by the same outlet (arxiv.org); overlapping topics (ability, context, model); pushes against this story (vs).
Reported by the same outlet (arxiv.org); overlapping topics (adaptive, agent, attack); pushes against this story (but).
Reported by the same outlet (arxiv.org); overlapping topics (attack, benign, model); pushes against this story (but).
Reported by the same outlet (arxiv.org); overlapping topics (attack, benign, context); pushes against this story (but).
Reported by the same outlet (arxiv.org); overlapping topics (agent, attack, model); pushes against this story (against).
Reported by the same outlet (arxiv.org); overlapping topics (agent, context, model); pushes against this story (against).