Skip to content

Scan record

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

Nov 1639 loaded

  1. The Right Tokens, the Right Time: From Per-Token Gating to Trajectory-Aware Activation Steering

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

  2. From Bytes to Behavior: Request-Level Attribution from Hidden States of Running GPUs

    PaperarXiv: 其他 AI 安全论文

  3. LLM puts the existential risk of AI at 10%

    PaperarXiv: 其他 AI 安全论文

  4. Forget the lineage: forward-secure, lineage-complete cryptographic erasure for LLM agent memory

    PaperarXiv: 其他 AI 安全论文

  5. Reasoning Models Confabulate Explanations to Defend Injected Beliefs About Their Own Requests

    PaperarXiv: 其他 AI 安全论文

  6. Adversarial attacks on symbolic vocabulary vulnerabilities in LLM planners

    PaperarXiv: 其他 AI 安全论文

  7. Cracks in the Embedding: Robustness and Stability of JEPA-Based Chest X-ray Representations

    PaperarXiv: 其他 AI 安全论文

  8. Soteria: Formally Verified Planning and Runtime Enforcement for Safe LLM Agents

    PaperarXiv: 其他 AI 安全论文

  9. Anti-distillation is cheap: lightweight distillation resistance via style disruption

    PaperarXiv: 其他 AI 安全论文

  10. Auditing Chatbots' Encroachment into the Professional Legal Practice Domain

    PaperarXiv: 其他 AI 安全论文

  11. Safety side effects of preventive personality steering: an empirical audit

    PaperarXiv: 其他 AI 安全论文

  12. System security is only as strong as the weakest agent: an attack that occurs only in multi-agent systems

    PaperarXiv: 其他 AI 安全论文

  13. Pushing Down the Mountain: The Effect of Framing Effects and Internal State Interventions on Collusive Behavior of Long-Horizon LLM Agents

    PaperarXiv: 其他 AI 安全论文

  14. Adversarial tracking: attacks on image sensors for autonomous vehicle object tracking

    PaperarXiv: 其他 AI 安全论文

  15. Unfinished-Bench: Measuring Rule-Violating Behavior of Persistent Agents in Real Work Environments

    PaperarXiv: 其他 AI 安全论文

  16. Physical Attacks on End-to-End Autoencoder-Based Communications via Evolutionary Transfer Learning

    PaperarXiv: 其他 AI 安全论文

  17. Google Project Zero will shorten vulnerability disclosure deadlines in stages: 60 days starting in 2027, 30 days starting in 2028

    NewsiThome (Taiwan)Governance

  18. New York Times: Lutnick family business's ties to Malaysian data centers raise conflict-of-interest questions

    NewsSeoul Economic DailyGovernance

  19. Cheating in China's college entrance exam can carry a maximum sentence of 7 years in prison; AI question-solving features suspended during exams

    NewsSeoul Economic DailyEvent

  20. Top scorer in South Korea's tax accountant exam admits to cheating with a miniature camera and generative AI and turns himself in

    NewsFinancial News KoreaEvent

  21. AI cheating suspicions spread, calls emerge in multiple South Korean national exams to verify high scorers

    NewsChosunBizEvent

  22. South Korean prime minister asks for study on shutting down Wi-Fi at CSAT exam venues to prevent AI cheating

    NewsYonhapEvent

  23. Indian officials say India is less concerned about AI threats than the United States

    NewsSemaforAI updates

  24. Taiwan's deputy digital minister on the Jensen Huang scandal: "We feel sad"

    NewsSemaforGovernance

  25. A generalizable safety-aware robot behavior planning method based on an explicit constraint space

    PaperarXiv

  26. Adaptable Watermarking: Transforming Image Watermarking Models via Lightweight Adapters

    PaperarXiv

  27. VLAN: Vision-Language Accessible Navigation

    PaperarXiv

  28. Securing Multi-Agent Systems by Design

    PaperarXiv

  29. An operational definition of opaque serial depth

    PaperarXiv

  30. Beyond Attack-Based Accuracy: Rethinking Adversarial Robustness Evaluation via Structural Gradient Alignment

    PaperarXiv

  31. Decoupling Forgetting and Utility: Lightweight LLM Unlearning via Orthogonal Null-Space Projection

    PaperarXiv: Jailbreaks, Prompt Injection, Poisoning & Defenses

  32. Has Abliteration Become Harder? Not Consistently: Direction-Based Ablation and ARA for Refusal Removal in Qwen3.8, Gemma 4, Olmo 3.1 and Mistral Small 3.2 LLMs and Their Predecessors in Heretic

    PaperarXiv: Jailbreaks, Prompt Injection, Poisoning & Defenses

  33. Constructing a steering vector via input optimization

    PaperarXiv: Jailbreaks, Prompt Injection, Poisoning & Defenses

  34. Alignment via Reflect-Reasoning: Learning How to Follow Principles

    PaperarXiv: Jailbreaks, Prompt Injection, Poisoning & Defenses

  35. FT Lex commentary: publishers are capable of a tough contest with big tech companies

    NewsFinancial Times · AIAI updates

  36. Former Nvidia chief Jensen Huang steps down in disgrace, dealing a blow to "open" models, with Anthropic and OpenAI benefiting

    NewsFinancial Times · AIAI updates

  37. Japan plans to regulate unauthorized commercial use of AI to replicate celebrities' voices and likenesses

    Newskmkj.krGovernance

  38. OpenAI Foundation board member says AI incidents will become the most likely cause of human death, warns of AI losing control

    NewsSegye IlboOpinion

  39. US Commerce Secretary Lutnick reportedly clears AI chips and advocates lax regulation, possibly involving family business interests in Southeast Asia

    Newsglobal.hk01.comGovernance