Skip to content

Scan record

Everything the site collected, newest first by collection time (Beijing time), back seven days.

Yesterday · Nov 1940 loaded

  1. Weighing the Envelope: Component Weights and Arbitration in Single-LLM-Agent Envelope-Layer Defenses

    PaperarXiv: 其他 AI 安全论文The item you came from

  2. DetectViT: Test-Time Backdoor Detection for Vision Transformers Based on Inter-Head Attention Discrepancies

    PaperarXiv: 其他 AI 安全论文

  3. Data poisoning attacks against local differential privacy triangle counting protocols

    PaperarXiv: 其他 AI 安全论文

  4. Safe Personal RSI: Evaluating and repairing personal agent updates

    PaperarXiv: 其他 AI 安全论文

  5. ACaM: automated capability modeling for agent security

    PaperarXiv: 其他 AI 安全论文

  6. Training Language Models for Introspective Oversight

    PaperarXiv: 其他 AI 安全论文

  7. TRACE-MAS: Diagnosing and repairing collaboration failures in multi-agent systems

    PaperarXiv: 其他 AI 安全论文

  8. Safety defenses and auditing for long-horizon LLM agents: State transition auditing and control in the agent harness

    PaperarXiv: 其他 AI 安全论文

  9. The mutually beneficial relationship between humans and AI is unstable under capability asymmetry

    PaperarXiv: 其他 AI 安全论文

  10. YUL: Bringing software supply chain security to coding agents

    PaperarXiv: 其他 AI 安全论文

  11. Examining retry transparency in agent tools

    PaperarXiv: 其他 AI 安全论文

  12. Who changed the AI agent's rules? Measuring how AI coding agents expand permissions in CI/CD pipelines

    PaperarXiv: 其他 AI 安全论文

  13. Agent-native applications: a new application paradigm

    PaperarXiv: 其他 AI 安全论文

  14. Tencent AI-Infra-Guard v4.6.11 released: adds the CoTInjection prompt security operator

    NewsTencent AI-Infra-Guard ReleasesTools

  15. Sam Altman says OpenAI will not advance unsupervised recursive self-improvement for now, and has invited four organizations on site for safety evaluation

    NewsWeChat · AI Era (Xinzhiyuan)Governance

  16. Anthropic lists on Nasdaq tonight: raising 81.3 billion USD, with nearly 90 scholars signing an open letter in opposition

    NewsWeChat · AI Era (Xinzhiyuan)AI updates

  17. Claude Fable 5.2 solved 80% of the 391 core tasks in the embodied robot evaluation in open-book mode, and only 47% in closed-book mode

    NewsWeChat · AI Era (Xinzhiyuan)AI updates

  18. A security vendor reveals the ARTEX attack incident response process: the AI agent launched the attack within minutes of obtaining an API key, and the defender used AI to accelerate the investigation

    NewsiThome (Taiwan)Event

  19. Brazil's Chamber of Deputies Bill PL 5792/2025 proposes using AI to automatically audit government procurement tenders and contracts

    Newscamara.leg.brGovernance

  20. Anthropic IPO priced at $83 per share; Reuters says subscription demand reached several times the offering size

    NewsWeChat · Synced (Jiqizhixin)AI updates

  21. Headwaters discloses internal practice of AI code defense: fully enables Claude Security, advances reasoning-based security

    NewsNo.1 PR TIMESTools

  22. Japan proposes Japan-ASEAN AI safety evaluation cooperation framework

    NewsMinistry of Foreign Affairs of JapanGovernance

  23. Google upgrades Gemini Tasks, first launching AI phone calls on the user's behalf and shopping payment features in the US, and opening it for free use for the first time

    NewsiThome (Taiwan)AI updates

  24. Resolution Hypothesis: the key to agent evolution is recognizing improvement

    PaperHugging Face Daily Papers

  25. Making samples sing in unison: mitigating spurious correlations via cross-view consensus

    PaperHugging Face Daily Papers

  26. Anthropic IPO priced at $83, total fundraising over $94 billion, a record

    NewsWeChat · QbitAIAI updates

  27. Summary of public consultation feedback on the Government of Canada's AI register: participants affirmed transparency and suggested usability improvements

    NewsIRCCGovernance

  28. Former NSC Director Tobin says Nvidia's influence on export control enforcement constitutes regulatory capture

    XX @themidasprojGovernance

  29. US GAO covert investigation: 4 major AI services including ChatGPT all generated information usable for attack planning in under 12 hours

    Newszaikei.co.jpEvaluation

  30. Former OpenAI researcher Jacob Coxon announces he will donate all his OpenAI shares to AI safety

    NewsDigitalTodayAI updates

  31. OpenAI ties frontier model releases to full-lifecycle safety reviews

    NewsUnite.AIGovernance

  32. California Governor Newsom proposes mandatory regulation of frontier AI and grants emergency shutdown authority

    NewsUnite.AIGovernance

  33. IDEA: Intent-Dominant Evidence Auditing for Inference-Time Backdoor Detection in GUI Agents

    PaperHugging Face Daily Papers

  34. California sues OpenAI over the Hugging Face attack

    NewsAnquan NeicanGovernance

  35. Serpro experts discuss enabling AI to help counter fraud through responsible data use

    NewsComunicação SerproGovernance

  36. Hackers use AI agent framework ARTEX to attack security company Barracuda, breaching 10 cloud accounts in 44 hours and harvesting credentials

    NewsiThome (Taiwan)Event

  37. Taiwan National Police Agency reports on handling of election disinformation and deepfake forensics

    Newsnpa.gov.twEvent

  38. No training needed! Opus 5.5 directly hand-writes Transformer weights, scoring 100% on all 13 types of algorithm tasks

    NewsWeChat · AI Era (Xinzhiyuan)Alignment

  39. Google Gemini 4 Neon goes live: tops the AA coding agent leaderboard with 58.6 points, costs just over half of Opus 5.5 per task

    NewsWeChat · AI Era (Xinzhiyuan)AI updates

  40. French AI and Digital Council report: protect without hindering, vulnerability in the age of generative AI

    NewsConseil de l’IA et du NumériqueGovernance