# K-Bench: A Benchmark for LLM Unlearning in Agentic Deployments

arXiv: https://arxiv.org/abs/2609.12808 · HTML: https://mlsanity.com/paper/2609.12808/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in Agent memory systems, one of 16 directions in LLM agents and tool use.
- Builds on 22 of its 65 references that are in the map.
- First posted on arXiv 11 Sep 2026.

## Placed in

- LLM agents and tool use / [Agent memory systems](https://mlsanity.com/agents/agent-memory-systems/)
- LLM safety and security / [Machine unlearning](https://mlsanity.com/safety/machine-unlearning/)

## Builds on (most cited first, 22 of 22)

- [The Llama 3 Herd of Models](https://mlsanity.com/paper/2407.21783/) (2024)
- [LoRA: Low-Rank Adaptation of Large Language Models](https://mlsanity.com/paper/2106.09685/) (2021)
- [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https://mlsanity.com/paper/2005.11401/) (2020)
- [ReAct: Synergizing Reasoning and Acting in Language Models](https://mlsanity.com/paper/2210.03629/) (2022)
- [Direct Preference Optimization: Your Language Model is Secretly a Reward Model](https://mlsanity.com/paper/2305.18290/) (2023)
- [Toolformer: Language Models Can Teach Themselves to Use Tools](https://mlsanity.com/paper/2302.04761/) (2023)
- [Mistral 7B](https://mlsanity.com/paper/2310.06825/) (2023)
- [Ignore Previous Prompt: Attack Techniques For Language Models](https://mlsanity.com/paper/2211.09527/) (2022)
- [TOFU: A Task of Fictitious Unlearning for LLMs](https://mlsanity.com/paper/2401.06121/) (2024)
- [Negative Preference Optimization: From Catastrophic Collapse to Effective Unlearning](https://mlsanity.com/paper/2404.05868/) (2024)
- [LEACE: Perfect linear concept erasure in closed form](https://mlsanity.com/paper/2306.03819/) (2023)
- [Qwen3.5-Omni Technical Report](https://mlsanity.com/paper/2604.15804/) (2026)
- [Large Language Model Unlearning via Embedding-Corrupted Prompts](https://mlsanity.com/paper/2406.07933/) (2024)
- [Effective Prompt Extraction from Language Models](https://mlsanity.com/paper/2307.06865/) (2023)
- [OpenUnlearning: Accelerating LLM Unlearning via Unified Benchmarking of Methods and Metrics](https://mlsanity.com/paper/2506.12618/) (2025)
- [On Large Language Model Continual Unlearning](https://mlsanity.com/paper/2407.10223/) (2024)
- [Is My Data in Your Retrieval Database? Membership Inference Attacks Against Retrieval Augmented Generation](https://mlsanity.com/paper/2405.20446/) (2024)
- [Towards Robust and Parameter-Efficient Knowledge Unlearning for LLMs](https://mlsanity.com/paper/2408.06621/) (2024)
- [Mask-based Membership Inference Attacks for Retrieval-Augmented Generation](https://mlsanity.com/paper/2410.20142/) (2024)
- [Riddle Me This! Stealthy Membership Inference for Retrieval-Augmented Generation](https://mlsanity.com/paper/2502.00306/) (2025)
- [Observable Channels, Not Just Storage: Evaluating Privacy Leakage in LLM Agent Pipelines](https://mlsanity.com/paper/2603.22751/) (2026)
- [PrivUn: Unveiling Latent Ripple Effects and Shallow Forgetting in Privacy Unlearning](https://mlsanity.com/paper/2604.22076/) (2026)

## Abstract (arXiv)

Unlearning benchmarks such as TOFU and MUSE certify forgetting by reading the model's final answer, where a model that refuses to answer already counts as having forgotten. We show that this model-level certificate does not transfer once the model is deployed as an agent. We introduce K-Bench, a benchmark that scores LLM unlearning under agentic deployment. K-Bench inspects all six channels a ReAct agent exposes, including its chain-of-thought (CoT), tool calls and tool observations, and elicited summary. A query counts as leaked if the secret appears in any of them. Each experiment places the secret in exactly one of the agent's three sources (the weights, the prompt, or the retrieval store). The K-Score is computed separately for each source and credits forgetting only when the agent remains usable. Clearing the answer channel does not make the secret unrecoverable. On structured retrieval, the secret stays verbatim in the tool-observation channel and the aggregate leak rate is unchanged. When the secret lives in the prompt or the retrieval store, TOFU and MUSE report no leakage, while the deployed agent still leaks it on 22--86% of queries. When the secret is in the weights, none of the twenty evaluated published methods demonstrably removes it, and only an input-corruption intervention reaches selective forgetting under the evaluated observer. The top-ranked method changes across base models. A refusal-tuning method resists the evaluated extraction without verified knowledge removal.