Anthropic이 평범한 하드웨어 질문에 대해 강화된 세이프티 필터를 적용하겠다고 협박함
Anthropic is threatening enhanced safety filters over completely normal hardware questions
핵심 요약
로컬 LLM 구동을 위한 하드웨어 사양을 묻는 평범한 질문에 Claude가 정책 위반 경고를 날리며 계정 제한을 예고해 논란입니다.
- 하드웨어 질문 검열 — 로컬 모델 구동용 RAM 및 서버 사양 문의가 정책 위반으로 오탐지됨
- 계정 제한 경고 — 반복적인 위반 시 강화된 세이프티 필터를 적용하겠다는 경고 메시지가 표시됨
- 필터 오작동 의혹 — AI 모델 학습(distillation) 금지 조항이 일반적인 하드웨어 문의와 혼동된 것으로 추측됨
- 대화 맥락 오염 — 한 번 플래그가 지정된 채팅 세션에서는 이후 모든 질문이 위반으로 처리되는 경향이 있음
TLDR:
- I asked Claude to research other people's experiences running large local models on a desktop with 256GB of RAM.
- The prompt was flagged before Claude replied.
- Claude still gave a completely normal hardware-related answer.
- I reported the apparent false positive using the contact Anthropic provided.
- I then asked whether an older high-memory server would be a cheaper alternative.
- That prompt was flagged as well.
- Anthropic is now warning that continued violations may result in enhanced safety filters being applied to my account.
I got this warning before Claude had even replied to my prompt:
It looks like a few of your recent prompts don’t meet our Usage Policy.
This is the chat where it happened:
https://claude.ai/share/a4285828-96d1-46ea-8dae-b147cd826694
My prompt was:
Are they right here?
I wouldn't hate upgrading to 256gb of ram to run models like glm5.2 @ 2bit with my 5090, i dunno if they're talking out of their ass or not (i have a 9800x3d) can you search what other peoples experiences with a build like that is
Claude then replied normally and discussed the hardware question. There was nothing unsafe or malicious in either my prompt or its response.
The warning linked to Anthropic's information about reporting issues with the safety system, so i emailed their User Safety address and included the chat link, the full prompt, and the screenshot i had sent Claude. I received what appeared to be an automated response directing me to the Safeguards Center.
A few minutes later, i sent another follow-up prompt in the same conversation:
hmmm i see, so even one of those old ddr3 servers on like this platform, is basically slightly worse and infinitely cheaper?
The prompt was accompanied by a product listing for a refurbished Dell PowerEdge R730 with two Xeon E5-2698 v3 CPUs and 256GB of RAM. I was comparing the cost and performance of an older server against upgrading my current desktop for local model inference.
