Activation steering offers a training-free path to defending large language models (LLMs) against jailbreak attacks, but existing methods fail to resolve the tension between safety and utility. We identify two root causes: safety concept vectors derived from difference-in-means are fragile under diverse attack geometries, and steering transformations lack a principled guarantee of leaving benign prompts unaffected. We propose AGOPNs (Average Gradient Outer Product with Null-space steering), which addresses both issues by extracting a safety concept vector via the Recursive Feature Machine, a feature-learning estimator grounded in the Average Gradient Outer Product, and confining the resulting steering matrix to the null space of benign activations, guaranteeing near-zero perturbation of benign prompts by construction. Experiments on Llama-3.1-8B-Instruct, Llama-3.1-70B-Instruct, and Llama-3.3-70B-Instruct across seven jailbreak attack families show that AGOPNs consistently achieves strong defense success rates while preserving near-baseline utility on AlpacaEval, XSTest, GSM8K, and MATH500.