Unverified50% confidenceCautionExact time
Soft tokens carry the risk of adversarial soft prompt attacks, where an attacker can craft vectors in the embedding space to manipulate the frozen LLM; norm clipping can serve as a mitigation measure but cannot fully eliminate the risk.
1
Sources
50%
Confidence
Long-term
Relevance
10/11/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedCALM performs minimal editing only on the token representations that are in violation, shifting them toward the safe side and suppressing residual components that are positively aligned with the unsafe direction.69% similarUnverifiedSilent patches are intended to reduce risk, but in practice may expand the attack surface by delaying users' proactive defenses.68% similarUnverifiedResearchers have found that fine-tuning only a small number of model parameters is sufficient to remove safety alignment (de-alignment attacks), and that certain jailbreak prompts can transfer across different models.65% similarUnverifiedSandbox isolation design can limit the impact scope when an agent encounters a malicious prompt injection attack65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/998490API
curl https://kongchang.com/api/v1/knowledge/claims/998490MCP
get_claim(id=998490)