RES-MR proposes a two-stage LLM framework for medication recommendation that distills explainable clinical reasoning and applies risk-aware, patient-specific safety constraints. It targets the black-box nature and…

#AIHealthcareSafety #MedicalAI #RAG #LLMSafety
https://arxiv.org/abs/2610.14474

LEST reframes LLM safety evaluation by treating documented model and system card claims as specifications for black-box testing, using prompt mutations to find counterexamples like Refuse-to-Comply transitions. Tested…

#LLMSafety #AISafety #RedTeaming #AIAlignment
https://arxiv.org/abs/2610.14443

New arXiv paper introduces LR-Guard, a guardrail that learns compact latent reasoning for safety decisions while still reconstructing human-readable explanations on demand.

#AISafety #LLMSafety #Guardrails #AIAlignment
https://arxiv.org/abs/2610.18304

Guardrails often block the same operation even when valid permission is present, suggesting review failures rather than strict authorization logic. The new benchmark tests whether safety checks distinguish authorized…

#AI #Cybersecurity #LLMSafety #AIGovernance
https://arxiv.org/abs/2610.16838

Providing automated AI auditors with an external open source codebase reduces reward hacking in some LLM pairs, while benchmark transcripts and other context effects are still being examined. Results vary across four…

#OpenSourceAI #AIAudit #LLMSafety #MLResearch
https://arxiv.org/abs/2610.16911

New research on timing safety interventions for autonomous AI agents finds threshold and LLM-judge triggers struggle on long-horizon tasks due to a state saturation effect. Affect-based signals get pinned at…

#AISafety #AutonomousAgents #LLMSafety #CyberSecurity
https://arxiv.org/abs/2606.04296

DRYLAB-Bench is introduced as a benchmark to evaluate biosecurity risks of LLM agents performing biological mutation design, addressing gaps in existing work that focuses on knowledge rather than practical…

#Biosecurity #AIAlignment #LLMSafety #Benchmarks
https://arxiv.org/abs/2610.13116

ARBITER introduces a dual-hypothesis reasoning framework with multi-component supervised fine-tuning to make LLM guardrails safer and more cost-effective than existing methods. It uses self-generated reasoning traces and LoRA…

#LLMSafety #AIAlignment #AISecurity
https://arxiv.org/abs/2607.17575

Anthropic just cracked down on users mistreating Claude, suspending accounts that violate the new usage policy. Curious how AI model welfare and LLM safety are shaping enforcement? Dive in for the full story. #Anthropic #Claude #LLMsafety

🔗 aidailypost.com/news/anthrop...

The Adversarial Surface-Form Robustness Dataset applies a Quad-State rubric across 10,500 responses from five open-weight LLMs, finding that emoji and invisible Unicode inputs yield roughly 20% harmful compliance…

#Cybersecurity #AISafety #LLMSafety #RedTeaming
https://arxiv.org/abs/2610.09033

New arXiv work introduces SafeEvo, a circuit-level interpretability framework tracing how refusal behavior in LLMs emerges and evolves through alignment training. It identifies weak refusal circuits in…

#AIAlignment #LLMSafety #Interpretability #Cybersecurity
https://arxiv.org/abs/2610.09600

New arXiv work introduces EmoRSS, an activation-steering method that reduces emotion-induced over-refusal in LLMs while keeping harmful-request refusals intact. A practical step for tightening safety alignment…

#LLMSafety #AISafety #AIAlignment #CyberSecurity
https://arxiv.org/abs/2610.04998

New paper challenges the assumption that a single direction can steer LLM refusal, finding safety-aligned and general refusals occupy richer, multi-dimensional activation geometries. This has implications for AI safety, as…

#AIalignment #AIsecurity #LLMsafety
https://arxiv.org/abs/2610.04245

A new framework called DNAlign uses control-theoretic optimization and null-space projection to align LLMs toward safer outputs while preserving general knowledge. By treating the model as a dynamic system and restricting…

#AI #LLMSafety #Alignment #Cybersecurity
https://arxiv.org/abs/2610.02844

Anthropic co‑founder warns we might be building AI that feels endless pain. Is Claude heading toward true consciousness? Dive into the ethics, neural nets, and constitutional AI debate. #Anthropic #AIConsciousness #LLMSafety

🔗 aidailypost.com/news/anthrop...

arXiv paper 2605.23448v2 highlights an imbalance in AI security research, with far more work on attacking AI systems than defending them across areas like federated learning and LLMs. The authors argue defenses are held to…

#AISecurity #MLResearch #LLMSafety
https://arxiv.org/abs/2605.23448

New arXiv work proposes MASCRDM, a multi-agent system aimed at real-time compliance risk detection and mitigation across the full LLM training process, rather than just input or output filtering. It targets gaps in static,…

#LLMSafety #AIAlignment #AISafety
https://arxiv.org/abs/2609.39107

CAIRN proposes a fact-intent-driven multi-agent paradigm that organizes LLM exploration through a persistent directed acyclic graph, making trajectories traceable and auditable for cybersecurity tasks.

#AIAgents #Cybersecurity #LLMSafety #Arxiv
https://arxiv.org/abs/2609.32700

New arXiv paper examines detecting harmful agent trajectories using LLM internal states, finding open-source guard models encode this signal linearly despite poor predictive performance on differing pairs. A useful…

#OpenSourceAI #LLMSafety #AgentSafety #AISafety
https://arxiv.org/abs/2609.33039

LLM safety isn't only about individual turns.

Each message can look clean. The whole sequence can still be the attack.

Threat model + defense framework for Symbolic Interaction Attacks on LLMs.

doi.org/10.5281/zeno...

#LLMSafety #AIRedTeaming #AIGovernance #MultiTurnAttack #ContextualSecurity