How to maintain model alignment after machine unlearning operations in text-to-image generative models.

Examining the OVERT benchmark that measures excessive refusal in text-to-image models, revealing how safety filters may be overly restrictive.

How execution-based verification reveals the extent and patterns of code hallucinations in large language models.

Investigating how the political leanings embedded in large language models can subtly influence voter behavior and democratic processes.

How TrojDiff demonstrates diverse trojan attack strategies against diffusion models, revealing vulnerabilities in generative AI systems.

Examining whether small-scale proxy training runs can reliably guide data curation decisions for large language model training.

How SafeKey leverages insight moments in chain-of-thought reasoning to improve LLM safety decision-making.

How evolving AI collectives can preserve and enhance human diversity while enabling self-regulation mechanisms in multi-agent systems.

How GuardAgent uses knowledge-enabled reasoning to build adaptive guardrails that protect LLM agents from adversarial exploitation.

Uncovering the mechanistic basis of LLM hallucinations through subsequence association analysis and what it means for building more reliable models.

A deep dive into agent runtime security and sandboxing solutions including OpenShell, OpenSandbox, CubeSandbox, Aegis EDR, and agent governance toolkits for safely running autonomous AI systems.

Exploring in-context watermarking techniques that embed detectable signals in LLM outputs without degrading generation quality.

The seminal work revealing that neural networks unintentionally memorize sensitive training data, enabling extraction of secrets like credit card numbers and passwords.

A comprehensive evaluation framework for measuring the trustworthiness and safety of video foundation models.

A breakthrough in provably undetectable watermarking for generative image models that preserves image quality while enabling reliable detection.

How situational context determines whether multimodal AI responses are safe or harmful, and a framework for context-aware safety evaluation.

The landmark demonstration that simple physical modifications to road signs can consistently fool deep learning-based visual classification in the real world.

A guide to emerging AI security threats including deepfake-as-a-service, the CLI-Anything agent backdoor attack, skill supply chain security, and the growing importance of content authenticity and provenance standards.

How SHINE provides provable defenses against backdoor attacks in deep reinforcement learning systems.

The pioneering work extending adversarial attacks beyond classifiers to generative models, revealing vulnerabilities in GANs and VAEs.

A critical analysis revealing that different notions of robustness in computer vision often conflict, and what this means for building truly reliable AI.

A position paper arguing that current approaches to AI agent security are fragmented and proposing a unified holistic framework for comprehensive agent safety.

How PixMix uses dreamlike image mixtures to simultaneously improve robustness, calibration, and safety across multiple computer vision benchmarks.

Extending training data extraction attacks to diffusion models, revealing that generative AI systems pose significant privacy risks through memorization.