defenses

This paper introduces GHOST, a failure mode where long-horizon LLM agents violate safety constraints specified many turns earlier under benign interactions, finding an 11.5% occurrence rate on GPT-5.5, and proposes STAR-Guard, a two-layer defense combining historical safety-constraint restoration with a pre-execution audit that eliminates GHOST events in experiments.

This paper argues that existing defenses against distillation attacks in large language models are ineffective when attackers further train their models with reinforcement learning, highlighting potential flaws in current threat models.

This paper provides a comprehensive survey of robustness in federated learning, covering threat models, aggregation strategies, defensive tactics, and future research directions.

Introduces AntiSkillBench, a benchmark for evaluating privacy leakage, impersonation risk, and defenses in persona-skill pipelines for AI agents, finding that risks persist across backbones and existing defenses are limited.

This paper investigates OS resilience against self-state attacks on self-hosted AI agents, characterizing an attack space and evaluating layered defense strategies. It finds that while a layered defense stack is effective, a small residual attack surface remains structurally indistinguishable at the OS level.

A discussion about handling prompt injection attacks in AI agents that read external content like emails and webpages, exploring production-level defenses and the subtle threats beyond obvious patterns.

Hacker Vitto Rivabella publicly announced the successful jailbreak of Fable 5, analyzing in detail the model's multi-layered security mechanisms, including input/output auditing, intent detection, and chain-of-thought defense, and provided methods to bypass them.