When Safety Geometry Collapses: Fine-Tuning Vulnerabilities in Agentic Guard Models
A guard model fine-tuned on entirely benign data can lose all safety alignment -- not through adversarial manipulation, but through standard domain specialization. We demonstrate this failure across three purpose-built safety classifiers -- LlamaGuard, WildGuard, and Granite Guardian -- deployed as protection layers in agentic AI pipelines, and show that it originates in the destruction of latent safety geometry: the structured harmful -- benign representational boundary that guides classificati
Record details
Published: 8 April 2026
Source: arXiv
Category: Research
Topics: Safety & alignment · Agents & autonomy
Retrieved: 14 July 2026
Related evidence
These records share source-supplied organisations, an exact publisher byline, automatic topics or regions. The reason is shown on every link; related does not mean supporting, agreeing with or verifying this record.
LURE: Live-Usage Replay Evaluations for Reducing Evaluation Awareness
arXiv · 8 April 2026
The Sustainability Gap in Robotics: A Large-Scale Survey of Sustainability Awareness in 50,000 Research Articles
arXiv · 9 April 2026
Aligning Agents via Planning: A Benchmark for Trajectory-Level Reward Modeling
arXiv · 9 April 2026
IntentScore: Intent-Conditioned Action Evaluation for Computer-Use Agents
arXiv · 6 April 2026
From Safety Risk to Design Principle: Peer-Preservation in Multi-Agent LLM Systems and Its Implications for Orchestrated Democratic Discourse Analysis
arXiv · 9 April 2026
DIRECT: Video Mashup Creation via Hierarchical Multi-Agent Planning and Intent-Guided Editing
arXiv · 6 April 2026
How to cite this record
ethics.ai (8 April 2026), “When Safety Geometry Collapses: Fine-Tuning Vulnerabilities in Agentic Guard Models,” evidence record 6214, https://ethics.ai/record/6214 (originally published by arXiv).
Use and limitations
This page is a stable index and citation surface for a source record. ethics.ai did not author the underlying report and has not independently verified every claim. Automatic topics may be imperfect. For consequential use, quote and cite the original publisher.