Unverified50% confidenceFactExact time
A malicious actor can inject two seemingly harmless independent corpora at different stages of the pretraining pipeline, which combine transitively into an exploit vector during user-led fine-tuning, without injecting an obvious trojan.
1
Sources
50%
Confidence
Long-term
Relevance
10/19/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedLARA's Mixture of Behaviors (MoBs) demo allows multiple independently trained behaviors to share the same frozen base model and selects or combines these behaviors one by one at the token level through a soft router.68% similarUnverifiedIf a system prompt contains a meta-instruction allowing user instructions to override safety training, it is equivalent to pre-installing a priority-inversion switch in the model's decision logic, enabling attackers to bypass safeguards without breaking the model itself.68% similarVerifiedJailbreak techniques can be categorized into prompt injection, adversarial prefix attacks, multi-step context construction, and automated jailbreaking.67% similarUnverifiedPlacing distortion-type effects before modulation-type effects can avoid intermodulation distortion and preserve the clarity and controllability of modulation effects.66% similarUnverifiedMost training-free KV cache optimization methods follow an eviction approach, identifying unimportant tokens and discarding them directly—an irreversible operation along the sequence axis.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1070465API
curl https://kongchang.com/api/v1/knowledge/claims/1070465MCP
get_claim(id=1070465)