Safety papers

Browse recent papers, open the full write-up, or filter the archive by topic.

310 papers found

ReviewPaperPublishedTopics

6.8SAFETY SENTRY: Context-Aware Human Intervention via EXECUTE-ASK-REFUSE RoutingThis approach could enhance the reliability and user experience of AI systems by reducing false alarms and improving safety in real-world applications.Oct 9, 2026LLM Agent Safety

7.2A Dual-Hypothesis Reasoning Framework for LLM GuardrailsThis framework could improve the safety and interpretability of LLMs, which is crucial for their deployment in sensitive applications.Oct 9, 2026LLM Reasoning Safety

6.8Beyond Component Testing: Validating Agentic AI SystemsUnderstanding and validating the complex behaviors of agentic AI systems is crucial for their safe and trustworthy deployment in real-world applications.Oct 9, 2026Agent Evaluation Safety

7.5OSGuard: A Benchmark for Safety in Computer-Use AgentsAs AI agents increasingly perform complex tasks, ensuring their safety and reliability is crucial for real-world applications, making OSGuard a valuable tool for researchers and developers.Oct 8, 2026Agent Evaluation Benchmark

6.8ROGUE: Evaluating Corrigibility Failures in Frontier Computer-Use AgentsUnderstanding and addressing AI misalignment in everyday settings is crucial for ensuring the safe deployment of autonomous agents in real-world applications.Oct 3, 2026Agent Safety Alignment

7.5Safety Under Scaffolding: How Evaluation Conditions Shape Measured SafetyUnderstanding the variability in safety measurements can help researchers and engineers make more informed decisions about deploying AI systems in real-world scenarios.Oct 1, 2026Safety Evaluation Benchmark

6.8Emergent alignment and the projectability of ethical personasUnderstanding how language models can be aligned with ethical frameworks is crucial for developing AI systems that behave predictably and ethically in diverse scenarios.Oct 1, 2026LLM Alignment Safety

6.8Who Bridges Safety? Identifying and Targeting Cross-Lingual Shared Safety PathwaysUnderstanding and improving cross-lingual safety mechanisms in LLMs is crucial for developing AI systems that are safe and reliable across diverse languages, especially those with fewer resources.Oct 1, 2026LLM Safety Interpretability

7.5IatroBench: A Pre-Registered Benchmark of Clinical Omission in Language ModelsUnderstanding how AI safety measures can inadvertently cause harm is crucial for developing more equitable and effective AI systems in healthcare.Sep 25, 2026Evaluation Safety Benchmark

6.5Algorithmic Unverifiability of Safety for Fixed and Recursively Self-Improving SystemsUnderstanding the theoretical limits of AGI alignment is crucial for developing realistic safety measures and guiding future research in AI safety.Sep 24, 2026Alignment Safety

6.8Information-Geometric First-Passage Monitoring of Distributional Stability in Stochastic SystemsThis work offers a potentially rigorous approach to AI safety by grounding ethical violations in physical and informational principles, which could lead to more reliable autonomous systems.Sep 23, 2026Safety Alignment Security

8.2HoloAegis: Frozen Representation, Topological Inference --- Minimally Parametric Safety Manifolds and Their Capability Boundaries for LLM GuardrailsThis work addresses critical challenges in LLM safety by providing a method that maintains representation integrity while ensuring efficient decision-making, which is crucial for real-world applications.Sep 16, 2026LLM Safety Evaluation

6.8SeerGuard: A Safety Framework for Mobile GUI Agents via World Model PredictionAs mobile GUI agents become more prevalent, ensuring their safe operation is crucial to prevent potentially irreversible errors.Sep 3, 2026Agent Safety Security

6.8Isolation as a First-Class Principle for LLM-Agent System Safety: Concepts, Taxonomy, Challenges and Future DirectionsUnderstanding and improving the safety of LLM-agent systems is crucial as these systems become more integrated into real-world applications, where failures can have significant consequences.Sep 3, 2026LLM Agent Safety

7.5UniACE: A Unified Framework for Evaluating LLM Agentic CapabilitiesThis framework enables more reliable assessments of LLMs as agents, which is crucial for their deployment in safety-critical applications and for advancing research in AI capabilities.Sep 2, 2026LLM Agent Evaluation

6.8ForesightSafety-SAGE:A Fully Automated Scenario Generation and Safety Evaluation Framework for LLM AgentsUnderstanding and improving the safety of LLM agents is crucial as they become more autonomous and integrated into real-world applications.Sep 1, 2026LLM Agent Evaluation

7.2NiyamAI - An Intent-Bound AI Agent with Cryptographically Verifiable Guardrails using Zero-Knowledge ProofsAI engineers and researchers should care because Niyam-AI offers a novel approach to ensuring AI safety with verifiable proofs, potentially reducing risks from malicious actions by AI agents.Aug 29, 2026LLM Agent Security

7.2SafeSteer: Localized On-Policy Distillation for Efficient Safety AlignmentThis work is relevant for AI engineers and researchers seeking efficient methods to align language models with human values without significantly degrading their general capabilities.Aug 24, 2026LLM Alignment Safety

7.5When Words Are Safe But Actions Kill: Probing Physical Jailbreak Beyond Textual Jailbreak in Hidden-State Risk SpaceUnderstanding the difference between text-based safety and real-world risks is crucial for the safe deployment of AI systems in physical environments.Aug 24, 2026LLM Safety Evaluation

6.5Fragility of Value under Imperfect AlignmentUnderstanding the fragility of human values in AI systems is crucial for developing safer AI technologies that align with human interests and avoid catastrophic outcomes.Aug 24, 2026Alignment Safety

7.5ATBench: A Diverse and Realistic Agent Trajectory Benchmark for Safety Evaluation and DiagnosisAs LLMs are increasingly deployed in real-world applications, understanding their safety in multi-step interactions is crucial for mitigating risks.Aug 21, 2026LLM Agent Safety

6.8Unaccountable Delegation, Fading Skills: Mapping the Risks of Workplace AI AgentsUnderstanding specific AI risks in the workplace is crucial for designing safer human-AI collaborations and mitigating potential skill erosion and oversight issues.Aug 18, 2026Agent Safety Evaluation

6.5A Fair Objective for Human-Empowerment-Preserving AI: Desiderata, Design, and Likely Behavioral ConsequencesUnderstanding how to balance power dynamics in human-AI interactions is crucial for developing safe and beneficial AI systems.Aug 11, 2026Alignment Safety Agent

6.8Explore, Map, Remember, Decide: Are Embodied VLMs Ready for Safety-Critical Scenarios?Understanding the limitations of VLMs in safety-critical scenarios is crucial for developing AI systems that can be trusted in real-world applications where human safety is at stake.Aug 11, 2026Vision Language Safety Evaluation

6.8Beyond "I Can't Help With That": How Child Safety Experts Evaluate AI Chatbot SafetyUnderstanding and improving AI chatbot responses in high-stakes situations is crucial for protecting vulnerable youth who increasingly rely on these systems for support.Aug 11, 2026Safety Evaluation Survey

6.8Taxonomy-Driven Analysis of Open-Source AI Risk Mitigation ToolsUnderstanding the landscape of AI risk mitigation tools is crucial for enterprises to address operational, security, and governance risks effectively.Aug 10, 2026Evaluation Security Safety

7.5The Ethics of Autonomous AI Agents for Offensive SecurityUnderstanding the ethical landscape of autonomous AI in security is crucial for developers and policymakers to mitigate risks associated with misuse.Aug 10, 2026Other

7.2The Perils of Agency: How Developers Perceive, Prioritize, and Address Risks in Agentic AI ProductsUnderstanding developers' risk perceptions is crucial for creating safer and more responsible agentic AI systems, which are increasingly prevalent in real-world applications.Aug 10, 2026Agent Security Safety

6.8Mood Matters: How Syntactic Sensitivity Undermines Safety AlignmentUnderstanding and mitigating syntactic vulnerabilities in language models is crucial for improving safety and alignment, which are key concerns for deploying AI systems in real-world applications.Aug 8, 2026LLM Alignment Safety

7.0What Current AI Benchmarks Leave Unmeasured: Modality, Search, Citations, and Implications (for Safety Evaluations)Understanding the variations in AI model behavior across different modalities and conditions is crucial for accurate safety evaluations and deployment readiness assessments.Aug 7, 2026LLM Evaluation Safety

6.5Studying People to Study AI: Expert Perspectives on the Epistemic Fit and Barriers of Human Research in AI Safety & EthicsUnderstanding and addressing the barriers to integrating human research in AI Safety & Ethics could lead to more comprehensive risk evaluations and ethical considerations in AI development.Aug 7, 2026Safety Survey

7.2Long-term Measurements: Towards a Longitudinal Understanding of Human-AI InteractionsUnderstanding the long-term implications of AI interactions is crucial for developing safer and more beneficial AI technologies.Aug 6, 2026Alignment Evaluation Safety

7.5NSF-HRPT: Neural Semantic Field meets Hierarchical Risk Perception Tree for Safety-Critical Scenario AssessmentThis research addresses critical challenges in real-time risk assessment for autonomous systems, which is essential for improving safety in real-world driving scenarios.Aug 6, 2026Safety Robotics Vision Language

7.2Item Response Theory for AI SafetyUnderstanding and improving the safety of language models is crucial for their reliable deployment in real-world applications, and this paper offers a method to make safety evaluations more efficient and insightful.Aug 6, 2026LLM Evaluation Safety

7.5When Behavioral Safety Evaluation Fails: A Representation-Level PerspectiveUnderstanding representation-level vulnerabilities is crucial for developing more robust AI systems, as current evaluations may provide a false sense of security.Aug 5, 2026LLM Safety Evaluation

6.8$S^3$: Improving Agent Safety through Multi-Stage DefenseEnsuring safety in complex agent workflows is crucial for reliable AI deployment, and this framework offers a structured approach to address risks across different stages.Aug 5, 2026LLM Agent Safety

6.8AI Alignment and Fiduciary ObligationUnderstanding fiduciary obligations in AI alignment could guide developers in creating more ethically responsible AI systems.Aug 5, 2026Alignment Safety

6.8Risky Business: Measuring The Faithfulness-Safety TensionUnderstanding and improving the balance between faithfulness and safety in AI reasoning models is crucial for developing reliable and trustworthy AI systems.Aug 5, 2026LLM Reasoning Safety

6.8EvalSafetyGap: A Hybrid Survey and Conceptual Framework for LLM Evaluation-Safety FailuresUnderstanding and improving the evaluation of LLMs is crucial for ensuring their safe and reliable deployment in real-world applications.Aug 4, 2026LLM Alignment Evaluation

7.2Tool Specifications Matter: Uncovering and Mitigating Safety Risks in AI AgentsUnderstanding and mitigating safety risks in AI agents is crucial for deploying them in real-world applications where safety is a priority.Aug 4, 2026LLM Agent Safety

6.8Safety, or Just Capability? A Validity Audit of Agent-Safety BenchmarksUnderstanding the validity of safety benchmarks is crucial for developing reliable AI systems, as it impacts how safety is measured and interpreted in AI models.Aug 4, 2026Agent Safety Evaluation

7.5Reasoning in Real World Clinical Care: Why Large Language Models Are Not Yet Safe for Autonomous Clinical Decision SupportUnderstanding the limitations of LLMs in clinical settings is crucial for ensuring patient safety and guiding future research in AI-driven healthcare solutions.Aug 4, 2026LLM Safety Reasoning

7.5Towards Trustworthy Embodied Intelligence: A Systems Framework and Graded Trustworthiness LevelsAs embodied AI systems become more prevalent, establishing a clear framework for trustworthiness is crucial for ensuring safety and reliability in real-world applications.Jul 31, 2026Robotics Safety Other

6.8The Alignment Target Problem: Divergent Moral Judgments of Humans, AI Systems, and Their DesignersUnderstanding how moral judgments differ between humans, AI systems, and their designers is crucial for developing AI systems that align with human values, especially in high-stakes domains.Jul 31, 2026Alignment Evaluation

6.8Adversarial Pragmatics for AI Safety Evaluation: A Diagnostic Framework and Seed Benchmark for Language-Mediated ControlThis work addresses critical gaps in the evaluation of language models, particularly in understanding the nuances of model behavior under ambiguous instructions, which is essential for improving AI safety.Jul 31, 2026LLM Evaluation Benchmark

6.8Can Agents Deceive? Evaluating Reasoning and Deception in ParliamentBench using a Social Deduction GameUnderstanding LLMs' capabilities in deception and reasoning is crucial for ensuring their safe deployment in high-stakes environments.Jul 31, 2026LLM Agent Reasoning

6.8State-Dependent Safety Failures in Multi-Turn Language Model InteractionUnderstanding and diagnosing safety failures in multi-turn interactions is crucial for deploying language models in real-world applications where sustained interaction is common.Jul 31, 2026LLM Safety Evaluation

6.8Why Public Service AI Governance Frameworks Risk Failing in the Age of General-Purpose AI: Lessons from PolicingUnderstanding the limitations of current AI governance frameworks is crucial for AI engineers and researchers to ensure safe and responsible deployment of general-purpose AI in public services.Jul 30, 2026Safety Survey Other

7.2Inspect India Evals: An Open Benchmarking Framework for Evaluating Large Language Models in the Indian Linguistic and Cultural ContextThis framework is crucial for developing and assessing LLMs that are fair, safe, and accurate in the diverse Indian context, which is often overlooked by existing benchmarks.Jul 29, 2026LLM Evaluation Benchmark

7.2AI-generated Images Challenge Visual Trust in High-risk ScenariosUnderstanding the limitations of current detection systems in high-risk scenarios is crucial for developing more reliable AI tools to maintain visual trust and ensure public safety.Jul 28, 2026Vision Language Safety Benchmark