EU Regulatory Changes

This publication introduces a new evaluation framework for AI safety, moving beyond simple success rates to incorporate cost-aware metrics for both offensive and defensive security agents. The pape...

This paper, published on arXiv, introduces a novel methodology for assessing AI safety that goes beyond traditional text-based content filtering. The authors propose a "hidden-state risk space" app...

This paper, published on arXiv on July 16, 2026, details a novel cybersecurity vulnerability targeting AI coding agents. The research demonstrates that malicious actors can embed hidden instruction...

This paper, published on arXiv, introduces a novel method for automatically generating "leakage contracts" for black-box CPUs without relying on pre-defined templates. A leakage contract formally s...

This publication introduces DataShield, a novel technical framework designed to detect risky or non-compliant data used in the fine-tuning of large language models. The method works by identifying ...

This paper, published on arXiv, proposes a new cryptographic protocol called Non-Forward Secure Aggregation (NFSA) that uses a two-layer secret sharing scheme to enable secure data aggregation with...

This publication from arXiv presents a research paper detailing a new, simplified method for generating adversarial attacks against vision-language AI models, such as those used in multimodal searc...

This paper, published on arXiv, introduces a new mathematical criterion for evaluating the stability of network flow watermarking techniques, specifically a method called Causal IPD-QIM. While not ...

This paper, published on arXiv, proposes a new defensive technique called Random Logit Scaling (RLS) designed to protect deep neural networks from black-box score-based adversarial attacks. These a...

This paper, published on arXiv, presents a new framework for understanding and managing vulnerabilities in the open-source software ecosystem, specifically within the context of AI safety. It propo...