Fetching from the wire…
01
02
03
04
05
06
07
08
Skills2026-08-07 · source-backed
SkillJack found safety detection on poisoned trajectories ran 98.5% but fell to 11.4% on skills extracted from those same trajectories, with 80% of skill-mediated attacks persisting after the original records were deleted. Distillation launders intent. Cleaning your trace store isn't remediation.
Each link below shares sources, entities, or timing with this story.
Tencent's SkillJack work (arXiv 2608.03509) shows self-evolving agents launder malicious intent during skill extraction. Attack success rates of 56.2% on SkillX and 89.2% on Anything2Skill, with 80% of implanted skills surviving deletion of the original poisoned records. If yo...
arXiv 2609.11699 starts from the finding that On-Policy Self-Distillation degrades hard-reasoning performance, because conditioning on ground-truth solutions produces artificially confident traces that suppress hedging and penalize the exploratory self-correction hard problems...
Uses conditional mutual information to measure distillation-relevant info leakage through API outputs. Learns a transformation that strips distillation info while preserving task accuracy. Timely given Anthropic's industrial-scale distillation disclosure. (arXiv)
This BlackboxNLP reproduction re-ran the original distillation-transmits-preferences experiments with new preference categories, a new task (chess move generation), Ministral8B, and an answer-space ablation. The original claims hold, but transmission strength varies widely acr...
arXiv 2608.26733 presents an execution-only attack that reconstructs a hosted agent skill without ever asking the victim to reveal it, submitting crafted but ordinary tasks whose results discriminate between candidate hidden behaviors. At the weakest access level, final respon...
SecOPD fine-tunes a defense using token-level feedback during on-policy distillation rather than the sequence-level signal prior work used. Against PISmith adaptive injections on Qwen3.6-27B it reports 9.0% attack success where Meta-SecAlign, the previous state of the art, sit...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.