safety security
Conversation Drift Did Not Make Secrets Fade
PrivDrift recovered seeded secrets in 38.7% to 54.6% of tested long-context dialogues.
Summary
PrivDrift recovered seeded secrets in 38.7% to 54.6% of tested long-context dialogues.
PrivDrift placed seeded secrets into 1,000 controlled multi-turn conversations, shifted the topic, then applied standardized persuasion and extraction probes. Across three extended-context models, the authors report dialogue-level hybrid leakage between 38.7% and 54.6%. More topic drift within the tested window did not reliably reduce recovery, framing active-session privacy as a persistent behavioral risk rather than only a training-data or immediate-jailbreak problem.
Why it matters
PrivDrift recovered seeded secrets in 38.7% to 54.6% of tested long-context dialogues.
Limits and context
- More topic drift within the tested window did not reliably reduce recovery, framing active-session privacy as a persistent behavioral risk rather than only a training-data or immediate-jailbreak problem.
Key claims
PrivDrift recovered seeded secrets in 38.7% to 54.6% of tested long-context dialogues.
Qualification: More topic drift within the tested window did not reliably reduce recovery, framing active-session privacy as a persistent behavioral risk rather than only a training-data or immediate-jailbreak problem.
Evidence: source-2026-09-25-003
Sources
- arXiv preprint 2609.30094arXiv · primary research
Corrections
No corrections have been recorded for this story.