EU Regulatory Changes

This paper, published on arXiv under the AI Safety framework, introduces a new cryptographic technique called "Function Privatization" designed for the local differential privacy model. The core ch...

This paper, published on arXiv in July 2026, introduces a novel technical approach called "On-Policy Distillation" for improving the safety of large language models (LLMs). Rather than retraining a...

This publication introduces MemSecBench, a new benchmark framework designed to systematically test and measure memory poisoning vulnerabilities in AI agents. Memory poisoning occurs when an attacke...

This paper, published on arXiv, presents a new benchmark called HoF-Bench, which demonstrates that open-source, non-frontier AI models can rediscover real-world, previously AI-discovered Common Vul...

A new research paper, AgentSnare, has been published on arXiv that introduces a framework for defending against autonomous penetration testing agents. This is not a regulatory change itself, but it...

This publication introduces InkShield, a novel technical framework designed to protect individuals from unauthorized handwriting mimicry by AI systems. The paper details a method that subtly alters...

A new preprint from arXiv, titled "What Does It Take to Detect an AI Agent? Minimal Feature Sets for Behavioral Detection under Browser Automation," published on July 29, 2026, presents a research ...

This paper, published on arXiv, proposes a new defensive technique called alignment checking for detecting backdoor attacks in federated learning systems. Backdoor attacks occur when malicious part...

A new preprint titled ToxScreen: Detecting Whether an LLM Has Been Poisoned has been published on arXiv, proposing a method to identify whether a large language model has been deliberately compromi...

This paper, published on arXiv, introduces a novel framework called "Pre-hoc Failure Risk Inference" for multi-agent AI systems. Rather than detecting failures after they occur, the framework aims ...