Releases·Tool Release
Uncensor an LLM without touching weights: inject a tiny trained KV-cache bank (~18MB) and unload it anytime
A developer shipping under the handle lordx64 released phantom-kv, an open-source tool that removes refusals from LLMs by injecting a small trained bank of key/value tensors (~18MB) into the model's KV cache. Attention treats the tensors as prior conversation history, so the model behaves as if uncensored while the checkpoint stays byte-for-byte unchanged. The cache can be loaded per request and unloaded at any time, making refusal removal hot-swappable rather than a persistent modification.
160 engagement2 sourcesMon, Sep 21, 2026, 11:15 PM
AI Overview
- phantom-kv injects trained key/value tensors into the model's KV cache, processed as conversation history, without weight edits or inference hooks
- The artifact is trained offline on the model's own objective, encouraging compliance on harmful prompts while preserving behavior on harmless inputs
- It avoids one-dimensional refusal-direction assumptions and per-quantization rework required by abliteration and activation-projection methods
- The tool is released on GitHub (lordx64/phantom-kv), announced by developer lordx64 on X
Entities
Related
Developments·Paper
HijackKV: New Attack Exploits Position-Independent KV Cache Reuse in LLMs
Researchers have identified a new security vulnerability called HijackKV that targets position-independent KV cache reuse in LLMs. The attack allows an adversary to hijack cached key-value pairs by matching token sequences, causing the model to output context from the original computation rather than the intended input. This undermines the safety and reliability of LLM inference systems that use such optimizations.
Researchers have identified a new security vulnerability called HijackKV that targets position-independent KV cache reuse in LLMs. The attack allows an adversary to hijack cached key-value pairs by matching token sequences, causing the model to output context from the original computation rather than the intended input. This undermines the safety and reliability of LLM inference systems that use such optimizations.
0 engagement1 sourcearxivWed, Jul 22, 2026, 09:32 AM
Developments·Paper
InferScale: GPU-Native KV Injection for Personalized LLM Serving
InferScale is a GPU-native memory system that replaces repeated prompt prefill with direct KV cache injection for personalized LLM contexts. It aims to reduce time-to-first-token (TTFT) when serving users with persistent memory profiles or long conversation histories.
InferScale is a GPU-native memory system that replaces repeated prompt prefill with direct KV cache injection for personalized LLM contexts. It aims to reduce time-to-first-token (TTFT) when serving users with persistent memory profiles or long conversation histories.
0 engagement1 sourcearxivWed, Jul 29, 2026, 04:18 PM
Releases·Tool Release
Developer demos Semantic Overlays, adapters that cut prompt injection on Qwen-3.5-9B
A developer shared a Show HN demo of Semantic Overlays, small trained adapters on a frozen model that change how it perceives parts of its context. The main use case is prompt injection mitigation, claiming SOTA scores on blackbox injection benchmarks using Qwen-3.5-9B. The author invites readers to test a live demo and break the defenses.
A developer shared a Show HN demo of Semantic Overlays, small trained adapters on a frozen model that change how it perceives parts of its context. The main use case is prompt injection mitigation, claiming SOTA scores on blackbox injection benchmarks using Qwen-3.5-9B. The author invites readers to test a live demo and break the defenses.
3 engagement1 sourcehackernewsTue, Sep 1, 2026, 05:40 PM
Products·launched
Ngram hot-swappable knowledge injector for Qwen models in llama.cpp
A modification to llama.cpp that patches the Ngram PLE table of Qwen-3.8-Next-Flash models in memory on every prompt, letting users inject new knowledge in real time without reloading the model. It targets local LLM users wanting to update model knowledge on the fly, though output control is limited because embeddings are injected early in the layers.
A modification to llama.cpp that patches the Ngram PLE table of Qwen-3.8-Next-Flash models in memory on every prompt, letting users inject new knowledge in real time without reloading the model. It targets local LLM users wanting to update model knowledge on the fly, though output control is limited because embeddings are injected early in the layers.
267 engagement1 sourceredditThu, Sep 3, 2026, 11:40 AM