mlsanityData to 7 Oct 2026

Capability-Gated Language Models: Security Composes, Utility Does Not

arXiv 2609.00445 ↗First posted 31 Aug 2026cs.CR cs.AI cs.LGcited by 0

On the map: LLM safety and security → Safety alignment, refusal and guardrails

Abstract (arXiv)

Deployed language model safeguards (safety fine-tuning, filtering, unlearning) vary by principal only outside the model weights: filters are reconfigured, tiers are multiplied, and artefacts are reissued; inside one set of weights every request meets the same model configuration. This motivates us to define capability-gated deployment: per-principal access control inside one set of weights, whose configurations form a lattice - meets accumulate a principal's restrictions and joins pool a coalition's reach. We instantiate it by sparse rank gating over an existing nested-factorisation mechanism, guide profile search with one-pass attribution, and read every result once from a pre-registered held-out split. Security approximately composes: provably exactly at meets under a monotone-elicitation assumption we falsify pointwise. In two lineages the median held-out meet deepens suppression; the one effect surviving correction strengthens it. Utility does not: individually harmless profiles can compose to retention and fluency damage, and no compositional bound exists.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Placed in Safety alignment, refusal and guardrails, one of 12 directions in LLM safety and security. Builds on 18 of its 30 references that are in the map. First posted on arXiv 31 Aug 2026.

Builds on

18 of 30 references.

Closest ancestors · 2

Same direction, newest first

  1. 2024The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearning5 Mar
  2. 2024Assessing the Brittleness of Safety Alignment via Pruning and Low-Rank Modifications7 Feb

Other references · 16

  1. 2022Training language models to follow instructions with human feedbackcited by 15,712
  2. 2023Direct Preference Optimization: Your Language Model is Secretly a Reward Modelcited by 7,916
  3. 2020Measuring Massive Multitask Language Understandingcited by 7,768
  4. 2022Locating and Editing Factual Associations in GPTcited by 2,609
  5. 2020The Pile: An 800GB Dataset of Diverse Text for Language Modelingcited by 2,346
  6. 2017Learning Important Features Through Propagating Activation Differencescited by 1,454
  7. 2019Designing and Interpreting Probes with Control Taskscited by 608
  8. 2019Importance Estimation for Neural Network Pruningcited by 531
  9. 2023LEACE: Perfect linear concept erasure in closed formcited by 254
  10. 2023Attribution Patching Outperforms Automated Circuit Discoverycited by 196
  11. 2025Measuring AI Ability to Complete Long Software Taskscited by 142
  12. 2024An Adversarial Perspective on Machine Unlearning for AI Safetycited by 113
4 more
  1. 2023The MiniPile Challenge for Data-Efficient Language Modelscited by 74
  2. 2024Do Unlearning Methods Remove Information from Language Model Weights?cited by 66
  3. 2025Deep Hierarchical Learning with Nested Subspace Networks for Large Language Modelscited by 5
  4. 2026No More, No Less: Least-Privilege Language Modelscited by 1

Built on by

Not cited yet.