EU Regulatory Changes

The publication introduces ColluSkill, a novel adversarial technique that demonstrates how malicious actors can evade AI agent safety scanners by composing multiple benign skills in sequence to exe...

A new academic paper, titled "Full-Key Recovery and Forgery from One MQOM v2.1 Signature," has been published on arXiv. The paper demonstrates a practical cryptographic attack against the MQOM v2.1...

A new research paper, published on arXiv, demonstrates that analyzing a large language model's internal activations can reveal whether it is generating insecure code, even when the model's final ou...

A new research paper, published on arXiv, challenges the reliability of internal harmfulness scores used to evaluate AI safety. The study demonstrates that these scores, which are often used to ran...

This paper, published on arXiv, presents an independent reproducibility study of large language model (LLM) and agent-driven tools designed to automatically validate software vulnerabilities. The s...

A new academic paper, published on arXiv, proposes a novel technical method called Dual-Adversarial Safety Alignment to improve the safety of large reasoning models (LRMs). The paper argues that cu...

The publication introduces RangeFactory, a technical framework for automatically generating multi-hop cyber range scenarios, which are realistic training environments for simulating complex cyberat...

A new research paper, titled STAIR, proposes an end-to-end agentic planning framework designed to improve incident response for AI systems. Published on arXiv, this is not a regulatory mandate but ...

A new academic paper, titled Sound Enforcement of Dynamic Release Information Flow Policy, has been published on arXiv. It proposes a technical framework for enforcing dynamic information flow cont...