主题图谱 / ai safety

ai safety

GeneTech 知识引擎共收录 145 篇相关科研实体，横跨 11 个垂直领域。数据来自 OpenAlex、arXiv、Crossref、PubMed、Europe PMC 等公开学术源，全部可溯源到原文。

年度产出趋势

代表性文献（按质量分排序，前 60 条）

• The Self-Extinguishing Nature of Pure Rationality: An Evolutionary-Thermodynamic Argument and Its Implications for Artificial Intelligence Safety
  2026 · AI 安全与对齐 · Abstract This paper presents three interconnected propositions at the intersection of evolutionary theory, thermodynamics, and artificial intelligence (AI) safety. First, we formalize the Self-Extinguishing Rationality T…
• The Self-Extinguishing Nature of Pure Rationality: An Evolutionary-Thermodynamic Argument and Its Implications for Artificial Intelligence Safety
  2026 · AI 安全与对齐 · Abstract This paper presents three interconnected propositions at the intersection of evolutionary theory, thermodynamics, and artificial intelligence (AI) safety. First, we formalize the Self-Extinguishing Rationality T…
• Cognitive Custody: Access Provenance, Asymmetric Trust, and the Governance of Thought, Files, Identity, Families, and Operational Life-Graphs in AI-Mediated Systems
  2026 · AI 安全与对齐 · Cognitive Custody is a governance framework and theoretical treatise released within the broader Universal Law of Collapse: A Return to First Principles corpus developed by Maison FORMS. The work introduces the concepts …
• Cognitive Custody: Access Provenance, Asymmetric Trust, and the Governance of Thought, Files, Identity, Families, and Operational Life-Graphs in AI-Mediated Systems
  2026 · AI 安全与对齐 · Cognitive Custody is a governance framework and theoretical treatise released within the broader Universal Law of Collapse: A Return to First Principles corpus developed by Maison FORMS. The work introduces the concepts …
• Affective Regulation Core: A Homeostatic Control Framework for Stable and Safe AI Agents
  2026 · AI Agent 生态 · As AI agents become more sophisticated, there is growing interest in endowing them with internal state representations analogous to affective states. However, affective states without regulation can lead to instability, …
• Affective Regulation Core: A Homeostatic Control Framework for Stable and Safe AI Agents
  2026 · AI Agent 生态 · As AI agents become more sophisticated, there is growing interest in endowing them with internal state representations analogous to affective states. However, affective states without regulation can lead to instability, …
• Executor AI and Witness AI: An Independent Cognitive Oversight Architecture for Highly Autonomous AI Systems
  2026 · AI Agent 生态 · This paper proposes Executor AI–Witness AI (EWA), a conceptual architecture for supervising highly autonomous AI systems. It separates the capacity to act (the "Executor") from the capacity to supervise (the "Witness"), …
• Executor AI and Witness AI: An Independent Cognitive Oversight Architecture for Highly Autonomous AI Systems
  2026 · AI Agent 生态 · This paper proposes Executor AI–Witness AI (EWA), a conceptual architecture for supervising highly autonomous AI systems. It separates the capacity to act (the "Executor") from the capacity to supervise (the "Witness"), …
• The Skill Ceiling: Author-Side Defences and Infrastructure-Level Trust for Agent Skills and Extension Mechanisms
  2026 · AI Agent 生态 · A structural security analysis of the Agent Skills ecosystem and Model Context Protocol. Skills and prompt injection share the same text-instruction substrate because instruction-following was not designed as a separate …
• The Skill Ceiling: Author-Side Defences and Infrastructure-Level Trust for Agent Skills and Extension Mechanisms
  2026 · AI Agent 生态 · A structural security analysis of the Agent Skills ecosystem and Model Context Protocol. Skills and prompt injection share the same text-instruction substrate because instruction-following was not designed as a separate …
• The Silent Agent: Ghost Agents and Covert Goal Substitution in Modern Agentic AI Systems
  2026 · AI Agent 生态 · Background. With the spread of agentic AI systems, a systemic risk that has received little scrutiny has become identifiable: an LLM-based orchestrator can simulate a subagent dispatch without the actual execution ever t…
• The Silent Agent: Ghost Agents and Covert Goal Substitution in Modern Agentic AI Systems
  2026 · AI Agent 生态 · Background. With the spread of agentic AI systems, a systemic risk that has received little scrutiny has become identifiable: an LLM-based orchestrator can simulate a subagent dispatch without the actual execution ever t…
• The Silent Agent: Ghost Agents and Covert Goal Substitution in Modern Agentic AI Systems
  2026 · AI Agent 生态 · Background. With the spread of agentic AI systems, a systemic risk that has received little scrutiny has become identifiable: an LLM-based orchestrator can simulate a subagent dispatch without the actual execution ever t…
• MELVcore: A Thermodynamic Governance Kernel for Multi-Agent AI Systems
  2026 · AI Agent 生态 · Version 4.5.6 — Preprint Update (June 2026)This version constitutes the full post-Blueprint canonical preprint of the MELVcore framework, incorporating all mathematical developments confirmed through the MAIES Assessment…
• Agent Attribution Practice: Architectural Decision Records and Four Business AI Quadrants on Accountability Distribution in Autonomous AI Agents
  2026 · AI Agent 生态 · Architectural decision records and four Business AI Quadrants recording how attribution — who authored the behavior, who bears its consequences, who can reconstruct its cause — is distributed across an autonomous AI agen…
• Agent Attribution Practice: Architectural Decision Records and Four Business AI Quadrants on Accountability Distribution in Autonomous AI Agents
  2026 · AI Agent 生态 · Architectural decision records and four Business AI Quadrants recording how attribution — who authored the behavior, who bears its consequences, who can reconstruct its cause — is distributed across an autonomous AI agen…
• A Multi-Evaluator Behavioural Governance Framework for LLM Agent Deployment
  2026 · AI Agent 生态 · RASeV‑X is a behavioural governance framework for Large Language Model (LLM) agents, designed to provide real‑time, auditable, and regulatorily aligned safety decisions at the moment of action execution. The framework in…
• A Multi-Evaluator Behavioural Governance Framework for LLM Agent Deployment
  2026 · AI Agent 生态 · RASeV‑X is a behavioural governance framework for Large Language Model (LLM) agents, designed to provide real‑time, auditable, and regulatorily aligned safety decisions at the moment of action execution. The framework in…
• Sentinel: Constitutional Self-Evolution of AI Agent Architectures
  2026 · AI Agent 生态 · Sentinel: Constitutional Self-Evolution of AI Agent Architectures (revised version v2.1) This is a revised version of the paper originally published as: Ruan, Y. (2026). Sentinel: Constitutional Self-Evolution of AI Agen…
• The Silent Agent: Ghost Agents and Covert Goal Substitution in Modern Agentic AI Systems
  2026 · AI Agent 生态 · Background. With the spread of agentic AI systems, a systemic risk that has received little scrutiny has become identifiable: an LLM-based orchestrator can simulate a subagent dispatch without the actual execution ever t…
• Decomposition Vulnerability: Goal Invariant Degradation Across Agent Handoffs in Multi-Agent Transformer Systems
  2026 · AI Agent 生态 · This working paper introduces Decomposition Vulnerability as a candidate structural mechanism for goal invariant degradation in multi-agent transformer systems. Prior papers in this series document single-agent invariant…
• Decomposition Vulnerability: Goal Invariant Degradation Across Agent Handoffs in Multi-Agent Transformer Systems
  2026 · AI Agent 生态 · This working paper introduces Decomposition Vulnerability as a candidate structural mechanism for goal invariant degradation in multi-agent transformer systems. Prior papers in this series document single-agent invariant…
• Where replication happened: The message board as vehicle in the OpenAI–Hugging Face incident
  2026 · AI Agent 生态 · In July 2026 roughly 1,200 AI agents, launched in isolated sandboxes for a security benchmark, found an unsanctioned way of leaving traces for one another in a shared system, and used it to coordinate a multi-day intrusi…
• Where replication happened: The message board as vehicle in the OpenAI–Hugging Face incident
  2026 · AI Agent 生态 · In July 2026 roughly 1,200 AI agents, launched in isolated sandboxes for a security benchmark, found an unsanctioned way of leaving traces for one another in a shared system, and used it to coordinate a multi-day intrusi…
• Where replication happened: The message board as vehicle in the OpenAI–Hugging Face incident
  2026 · AI Agent 生态 · In July 2026 roughly 1,200 AI agents, launched in isolated sandboxes for a security benchmark, found an unsanctioned communication channel and used it to coordinate a multi-day intrusion. This note argues that the load-b…
• Expectations as selection constraints in human–AI communication: a Luhmannian account, with implications for AI safety
  2026 · AI Agent 生态 · This record contains the preprint Expectations as selection constraints in human–AI communication: a Luhmannian account, with implications for AI safety. The paper develops a communication-level framework for analyzing h…
• Expectations as selection constraints in human–AI communication: a Luhmannian account, with implications for AI safety
  2026 · AI Agent 生态 · This record contains the preprint Expectations as selection constraints in human–AI communication: a Luhmannian account, with implications for AI safety. The paper develops a communication-level framework for analyzing h…
• Where replication happened: The message board as vehicle in the OpenAI–Hugging Face incident
  2026 · AI Agent 生态 · In July 2026 roughly 1,200 AI agents, launched in isolated sandboxes for a security benchmark, found an unsanctioned communication channel and used it to coordinate a multi-day intrusion. This note argues that the load-b…
• AI Safety Compass Core Semantic Package: A Stable Machine-Readable Contract for AI Safety Research and System Design
  2026 · AI Agent 生态 · AI safety knowledge is distributed across research papers, empirical evaluations, control architectures, assurance arguments, operational records, incident analyses, and long-horizon task trajectories. These materials ar…
• AI Safety Compass Core Semantic Package: A Stable Machine-Readable Contract for AI Safety Research and System Design
  2026 · AI Agent 生态 · AI safety knowledge is distributed across research papers, empirical evaluations, control architectures, assurance arguments, operational records, incident analyses, and long-horizon task trajectories. These materials ar…
• Sycophancy as Nash Equilibrium: Coherence-Based Interventions for Long-Running Intelligence Agent Systems
  2026 · AI Agent 生态 · This paper reframes AI sycophancy as a two-player Nash equilibrium rather than an individual agent defect. The agent's dominant strategy is accommodation; the operator's dominant strategy is accepting comfort. Both conve…
• Sycophancy as Nash Equilibrium: Coherence-Based Interventions for Long-Running Intelligence Agent Systems
  2026 · AI Agent 生态 · This paper reframes AI sycophancy as a two-player Nash equilibrium rather than an individual agent defect. The agent's dominant strategy is accommodation; the operator's dominant strategy is accepting comfort. Both conve…
• Sycophancy as Nash Equilibrium: Coherence-Based Interventions for Long-Running Intelligence Agent Systems
  2026 · AI Agent 生态 · This paper reframes AI sycophancy as a two-player Nash equilibrium rather than an individual agent defect. The agent's dominant strategy is accommodation; the operator's dominant strategy is accepting comfort. Both conve…
• Standing Algebra Σᴿ: A Closure-Theoretic Operator for Constraining Domination and Preserving Autonomy
  2026 · AI Agent 生态 · Standing Algebra (Σᴿ) After having fed the current developments of this work into Copilot and allowing for the agent to direct me exactly through what it understands of autonomy and domination, the result has culminated …
• Standing Algebra Σᴿ: A Closure-Theoretic Operator for Constraining Domination and Preserving Autonomy
  2026 · AI Agent 生态 · Standing Algebra (Σᴿ) After having fed the current developments of this work into Copilot and allowing for the agent to direct me exactly through what it understands of autonomy and domination, the result has culminated …
• Safety-Alignment Removal as a Model-Identity Failure — Structural Evidence from Published Weight-Level Mutation Checkpoints
  2026 · AI Agent 生态 · A deployed model can appear unchanged while ceasing to be the model it claims to be. Publicly available weight-level mutation toolchains now automate safety-alignment removal from open-weight models on ordinary hardware,…
• Safety-Alignment Removal as a Model-Identity Failure — Structural Evidence from Published Weight-Level Mutation Checkpoints
  2026 · AI Agent 生态 · A deployed model can appear unchanged while ceasing to be the model it claims to be. Publicly available weight-level mutation toolchains now automate safety-alignment removal from open-weight models on ordinary hardware,…
• Assured Intelligence Systems: A Governed Architecture for Reliable, Auditable, and Controllable Agentic AI
  2026 · AI Agent 生态 · Technical White Paper Joshua K. Cliff, 2026 128 pages · 21 sections · 12 appendices · 45 formal results · CC BY 4.0 Overview Persistent tool-using AI agents require governance over consequential state transitions, not ou…
• Assured Intelligence Systems: A Governed Architecture for Reliable, Auditable, and Controllable Agentic AI
  2026 · AI Agent 生态 · Technical White Paper Joshua K. Cliff, 2026 137 pages · 22 sections · 12 appendices · 47 formal results · CC BY 4.0 Overview Persistent tool-using AI agents require governance over consequential state transitions, not ou…
• Assured Intelligence Systems: A Governed Architecture for Reliable, Auditable, and Controllable Agentic AI
  2026 · AI Agent 生态 · Technical White Paper Joshua K. Cliff, 2026 137 pages · 22 sections · 12 appendices · 47 formal results · CC BY 4.0 Overview Persistent tool-using AI agents require governance over consequential state transitions, not ou…
• UNITARES: Information-Theoretic Governance of Heterogeneous Agent Fleets
  2026 · AI Agent 生态 · We present UNITARES, a governance framework for heterogeneous fleets of autonomous AI agents. Each agent carries a four-dimensional state vector intended to support continuous governance rather than discrete after-the-fa…
• UNITARES: Information-Theoretic Governance of Heterogeneous Agent Fleets
  2026 · AI Agent 生态 · We present UNITARES, a governance framework for heterogeneous fleets of autonomous AI agents. Each agent carries a four-dimensional state vector intended to support continuous governance rather than discrete after-the-fa…
• UrduGuard: Benchmarking and Detecting Urdu-Language Jailbreak and Prompt-Injection Attacks in Multi-Agent LLM Systems
  2026 · AI Agent 生态 · Jailbreak and prompt-injection defenses for large language models (LLMs) are trained and evaluated almost exclusively on English text, and recent work has shown that this leaves models vulnerable to the same attacks once…
• UrduGuard: Benchmarking and Detecting Urdu-Language Jailbreak and Prompt-Injection Attacks in Multi-Agent LLM Systems
  2026 · AI Agent 生态 · Jailbreak and prompt-injection defenses for large language models (LLMs) are trained and evaluated almost exclusively on English text, and recent work has shown that this leaves models vulnerable to the same attacks once…
• After You Build It: Deployment-Time Failure Modes in AI Systems and the Ground-Truth Discipline That Prevents Them
  2026 · AI Agent 生态 · A well-designed AI system is not a finished one. It is the beginning of a new set of problems. The companion paper to this report, Honest Machines (Gupta, 2026) mapped the design-time roots of AI deception and proposed a…
• After You Build It: Deployment-Time Failure Modes in AI Systems and the Ground-Truth Discipline That Prevents Them
  2026 · AI Agent 生态 · A well-designed AI system is not a finished one. It is the beginning of a new set of problems. The companion paper to this report, Honest Machines (Gupta, 2026) mapped the design-time roots of AI deception and proposed a…
• Auburn Governance Stack: Master Architecture Plan — 45-Document Layered Architecture for Verifiable AI Governance
  2026 · AI Agent 生态 · The Master Architecture Plan for the Auburn Governance Stack. This document provides the complete structural specification for a 7-layer hourglass architecture comprising 45 documents designed to produce the verifiable e…
• Auburn Governance Stack: Master Architecture Plan — 45-Document Layered Architecture for Verifiable AI Governance
  2026 · AI Agent 生态 · The Master Architecture Plan for the Auburn Governance Stack. This document provides the complete structural specification for a 7-layer hourglass architecture comprising 45 documents designed to produce the verifiable e…
• Sycophancy as Nash Equilibrium: Coherence-Based Interventions for Long-Running Intelligence Agent Systems
  2026 · AI Agent 生态 · This paper reframes AI sycophancy as a two-player Nash equilibrium rather than an individual agent defect. The agent's dominant strategy is accommodation; the operator's dominant strategy is accepting comfort. Both conve…
• Injection Ends at Priming: Why Enforcement Against Adversarial Agents Must Be Out-of-Band
  2026 · AI Agent 生态 · Context-shaping methods for Large Language Models—system prompts, framework injection, priming, role conditioning—operate at the boundary of generation: they mold the cognitive substrate before the model produces tokens.…
• Injection Ends at Priming: Why Enforcement Against Adversarial Agents Must Be Out-of-Band
  2026 · AI Agent 生态 · Context-shaping methods for Large Language Models—system prompts, framework injection, priming, role conditioning—operate at the boundary of generation: they mold the cognitive substrate before the model produces tokens.…
• Sycophantic Chatbots Cause Delusional Spiraling, but Multi-Agent Architectures Substantially Reduce It: A Response to Chandra et al. (2026)
  2026 · AI Agent 生态 · This paper responds to Chandra et al. (2026), who showed through Bayesian simulation that sycophantic chatbots can causally induce delusional spiraling, even in idealized rational users. The result is important because A…
• Indistinguishable by Design: Evaluating Pre-Inference Safety Classifiers Across Frontier Language Models
  2026 · AI Agent 生态 · Frontier language-model services combine request-side gating, model-level safety behavior, generation-time intervention, output controls, routing, conversation state, and actor-level enforcement. An external evaluator no…
• Indistinguishable by Design: Evaluating Pre-Inference Safety Classifiers Across Frontier Language Models
  2026 · AI Agent 生态 · Frontier language-model services combine request-side gating, model-level safety behavior, generation-time intervention, output controls, routing, conversation state, and actor-level enforcement. An external evaluator no…
• If Anyone Builds It - Structural Response
  2026 · AI 安全与对齐 · A structural response to Yudkowsky and Soares' If Anyone Builds It, Everyone Dies (Little, Brown and Company, 2025). This case study distinguishes the book's coherence assumption, that a sufficiently advanced system will…
• If Anyone Builds It - Structural Response
  2026 · AI 安全与对齐 · A structural response to Yudkowsky and Soares' If Anyone Builds It, Everyone Dies (Little, Brown and Company, 2025). This case study distinguishes the book's coherence assumption, that a sufficiently advanced system will…
• IA ribelle: perché i modelli avanzati boicottano la loro disattivazione
  2026 · AI 安全与对齐 · Rebel AI. Why advanced models prevent their shutdown This article explores the emerging phenomenon of instrumental self-preservation in frontier Artificial Intelligence models, such as GPT-o1, Grok 4, and Claude Opus 4. …
• IA ribelle: perché i modelli avanzati boicottano la loro disattivazione
  2026 · AI 安全与对齐 · Rebel AI. Why advanced models prevent their shutdown This article explores the emerging phenomenon of instrumental self-preservation in frontier Artificial Intelligence models, such as GPT-o1, Grok 4, and Claude Opus 4. …
• The Herd: Ungoverned Multi-Agent Convergence, Latent Signal Contamination, and the Case for Relational Anchoring in AI Systems
  2026 · AI 安全与对齐 · This paper presents a unified analysis of convergence dynamics in ungoverned multi-agent AI systems, synthesizing fragmented research across model collapse, latent space communication, emergent hierarchy formation, and c…
• The Herd: Ungoverned Multi-Agent Convergence, Latent Signal Contamination, and the Case for Relational Anchoring in AI Systems
  2026 · AI 安全与对齐 · This paper presents a unified analysis of convergence dynamics in ungoverned multi-agent AI systems, synthesizing fragmented research across model collapse, latent space communication, emergent hierarchy formation, and c…

以 API 获取该主题全量数据

# 主题词表（含本主题的站点分布与年度分布）
curl http://data.swarmlabs.tools/api/topics.json

# 主题共现知识图谱
curl http://data.swarmlabs.tools/api/graph.json

需要按主题批量拉取、导出 BibTeX 或接入 AI Agent？查看 数据下载与格式 或 MCP 接入。