2026-10-11 17:09 UTC

model-security

band: coolmomentum: stable score: 0.032
temperature history

Episodes (5)

Independent replication will determine whether previous-token prediction can reconstruct hidden LLM prompts with near-exact fidelity and create a practical prompt-confidentiality risk.
expiredconvergesscott: medium
Independent testing will determine whether Steganeur's token-choice encoding provides a reliable and practically usable covert channel in natural-looking LLM-generated text.
expiredconvergesscott: medium
The paper’s author claims constraining fine-tuning to subspaces learned from trusted LoRA adapters can block malicious model updates while preserving useful adaptation, potentially adding a geometric defense against fine-tuning poisoning.
expiredconvergesscott: medium
The authors of “Stealing Reasoning Traces from Proprietary LLM APIs” claim reasoning traces can be extracted from proprietary model APIs, potentially undermining providers' ability to keep those traces private.
expiredknownscott: low
OpenAI says it disrupted a coordinated model-distillation campaign attributed to Kimi, and Kimi's response plus any further disclosures or enforcement would establish organized cross-lab distillation theft as a recognized, actively policed frontier-model threat.
resolvedconvergesscott: high

Trajectory notes