phantom-kv: uncensor an LLM by injecting an ~18MB KV-cache bank, no weights touched
Anony6666 · reddit · 2026-09-22
A developer released phantom-kv, a refusal-removal system that modifies no model weights. It trains a compact bank of key/value tensors offline using the model's own objective, then injects it into the KV cache at inference, where attention treats it as conversation history — making the model compliant on targeted harmful prompts while preserving normal behavior on harmless inputs. Unlike weight-space abliteration (rewrites checkpoints, must be redone per quantization) or activation projection (engine-level hooks, assumes a single refusal direction), phantom-kv is a per-request, hot-swappable mode: unload it and the model remains byte-for-byte unchanged. Open-sourced at github.com/lordx64/phantom-kv. Notably, this demonstrates how fragile safety alignment can be against cache-level injection.
Related event: phantom-kv removes LLM refusals via 18MB KV cache injection(3 posts)→
More from Models
- Datalab claims parsing accuracy second only to Gemini at 1/37 the cost and 6x speed — VikParuchuri · 2026-11-12
- Datalab fast-mode OCR beats Opus 5.5 and GPT-6.1 Sol at 1/4 to 1/68 the price — VikParuchuri · 2026-11-12
- Gemini 4 Carbon opens to all at $2/$10, price doubling to $4/$20 on Dec 1 — socialwithaayan · 2026-11-12
- Ten wild Gemini 4 Carbon demos in 24 hours: CUDA meadows, V8 engines, mech stomps — socialwithaayan · 2026-11-12
- How to cut Fable 5.2 token costs by 90%: Arc routing plus a lean CLAUDE.md — EXM7777 · 2026-11-12
- Grok Bot 1.0 launches with support for joining Meet, Zoom and Teams calls — mark_k · 2026-11-12