Large Language Models Generate Harmful Content Using a Distinct, Unified Mechanism

This project represents a critical advance in Mechanistic Interpretability (MechInterp) applied to AI Safety. The core claim—that harmfulness is mediated by a 'unified mechanism' rather than being a diffuse property—is a high-stakes scientific hypothesis. If true, it allows for 'surgical' alignment or un-alignment. Quantitative signals (0 stars but 7 forks in just 7 days) are highly indicative of professional researcher interest; forks without stars often imply that practitioners are immediately cloning the code to run experiments rather than 'bookmarking' it for later. Defensibility is low (4) because the value lies in the scientific discovery and the methodology, which are easily replicated once published. There is no 'moat' in the code itself, as it uses standard PyTorch and MechInterp hooks. Frontier risk is 'high' because labs like Anthropic (pioneers of Dictionary Learning) and OpenAI (Safety Systems team) are the primary consumers and potential absorbers of this tech. If this mechanism is validated, frontier labs will immediately integrate it into their post-training/RLHF pipelines to harden models against jailbreaks or 'emergent misalignment.' Direct competitors include projects like 'The Refusal Vector' (Arditi et al.) and tools like TransformerLens. The displacement horizon is short (6 months) because the safety research cycle moves exceptionally fast, and labs will likely automate this type of pruning-based safety check within one or two training iterations.