AI ethics evidence index Search
Source-linked daily record
Evidence record 4847 · automatically gathered

Safety Anchor: Defending Harmful Fine-tuning via Geometric Bottlenecks

The safety alignment of Large Language Models (LLMs) remains vulnerable to Harmful Fine-tuning (HFT). While existing defenses impose constraints on parameters, gradients, or internal representations, we observe that they can be effectively circumvented under persistent HFT. Our analysis traces this failure to the inherent redundancy of the high-dimensional parameter space: attackers exploit optimization trajectories that are orthogonal to defense constraints to restore harmful capabilities while

Record details

Published: 7 May 2026
Source: arXiv
Category: Research
Topics: Safety & alignment · Military & security
Retrieved: 14 July 2026

source-onlyevidence status

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.

Shared topic: Safety & alignment · Shared topic: Military & security

One Turn Too Late: Response-Aware Defense Against Hidden Malicious Intent in Multi-Turn Dialogue

arXiv · 7 May 2026

Shared topic: Safety & alignment · Shared topic: Military & security

Latent Personality Alignment: Improving Harmlessness Without Mentioning Harms

arXiv · 8 May 2026

Shared topic: Safety & alignment · Shared topic: Military & security

Disentangling Intent from Role: Adversarial Self-Play for Persona-Invariant Safety Alignment

arXiv · 3 May 2026

Shared topic: Safety & alignment · Shared topic: Military & security

Acceptance Cards:A Four-Diagnostic Standard for Safe Fine-Tuning Defense Claims

arXiv · 11 May 2026

Shared topic: Safety & alignment · Shared topic: Military & security

SafeSteer: A Decoding-level Defense Mechanism for Multimodal Large Language Models

arXiv · 12 May 2026

Shared topic: Safety & alignment · Shared topic: Military & security

Fusion-fission forecasts when AI will shift to undesirable behavior

arXiv · 14 May 2026

How to cite this record

ethics.ai (7 May 2026), “Safety Anchor: Defending Harmful Fine-tuning via Geometric Bottlenecks,” evidence record 4847, https://ethics.ai/record/4847 (originally published by arXiv).

JSON

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.