SCOPE: Streaming Covariance-Orthogonal Post-Hoc Editing for Continual LLM Safety Governance

Deployed large language models (LLMs) require continual post-hoc safety updates to counter evolving threats, yet current methods often entangle refusal behaviors with capability-critical representations, causing cumulative utility drift in sequential edits. We propose SCOPE, a streaming post-hoc editing framework for capability-invariant LLM safety governance. SCOPE derives layer-wise capability subspaces via online activation covariance and constrains safety updates to their orthogonal spaces, decoupling capability preservation from safety-specific edits. Within the derived null space, it extracts low-rank refusal directions from contrastive harmful activations to steer refusal behaviors without interfering with core capabilities. Its online covariance maintenance enables efficient, adaptive expansion of the protected capability subspace against open-ended threat evolution. Our experiments validate that SCOPE delivers significant safety robustness gains by substantially reducing attack success rates, while keeping general capabilities within natural run-to-run variance under repeated edits, supporting long-term safety maintenance for deployed LLMs.