Skip to content

Scan record

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

Yesterday · Nov 1940 loaded

  1. LLM jailbreak detection based on contrastive prototype structure

    PaperarXiv: Jailbreaks, Prompt Injection, Poisoning & DefensesThe item you came from

  2. George Monbiot commentary: The UK growth plan forces through data centers nobody wants, betting that the AI bubble won't burst

    NewsThe Guardian · AIGovernance

  3. Researchers propose inducing monitorable reward hacking to compensate for the failure of gradient routing

    NewsLessWrongAlignment

  4. A LessWrong author posted about protesting Anthropic's IPO on November 19, calling for a ban on recursive self-improvement

    NewsLessWrongGovernance

  5. Shanghai AI Lab paper: switch to a different agent execution framework, and model score rankings may reverse

    NewsWeChat · Weijiqi LixinAI updates

  6. Tencent Zhuque Lab open-sources the pikit toolbox to systematically evaluate AI agent indirect prompt injection risk

    NewsWeChat · Weijiqi LixinTools

  7. Beihang and Alibaba teams propose the PASP framework, using multi-perception views and adversarial self-play to improve large models' ability to recognize disguised harmful content

    NewsWeChat · Weijiqi LixinDefenses

  8. Paper: under ANCHOR multi-turn attacks, refusal rates of 8 CLI agents drop to 0%

    NewsWeChat · Weijiqi LixinAttacks

  9. Xint discloses 5 Nginx RCE vulnerabilities, four discovered by automation

    NewsWeChat · SecureNexusLabEvent

  10. OpenAI is testing giving its autonomous agent Dots its own email address

    NewsiThome (Taiwan)AI updates

  11. Solomonoff induction based on reflective oracles

    NewsAI Alignment ForumAlignment

  12. CCN updates ENS information security policy and responsibilities and functions guidance, adding an appendix linking data protection and AI rules

    NewsCentro Criptológico NacionalGovernance

  13. Can you make a model believe false facts in the prompt?

    PaperarXiv: Alignment, Deception & Oversight

  14. Dependency fabrication in repository deployment agents

    PaperarXiv: Alignment, Deception & Oversight

  15. Nudge Security report: AI agents inherit employee access permissions, and their use is unsupervised

    NewsHelp Net Security AIEvent

  16. Negotiating control and accountability among stakeholders in high-risk AI application scenarios: development and evaluation of the ALIGN method

    PaperarXiv: 其他 AI 安全论文

  17. RECLAIM: robot vision-guided adversarial patch defense via embodied attention feedback

    PaperarXiv: 其他 AI 安全论文

  18. Truthgate: Checking coding agents' behavior against records they cannot edit—cheating, deterrence and disclosure on impossible tasks

    PaperarXiv: 其他 AI 安全论文

  19. PGPO: prototype-guided progressive obfuscation for privacy-preserving LLM-enhanced recommendation

    PaperarXiv: 其他 AI 安全论文

  20. Kaer: safety-aware replanning for multi-agent systems

    PaperarXiv: 其他 AI 安全论文

  21. Exactly stealing other clients' training data in federated learning

    PaperarXiv: 其他 AI 安全论文

  22. The limitation is in the input, not the model: state-dependent command safety in natural language interfaces for industrial controllers

    PaperarXiv: 其他 AI 安全论文

  23. Gradient-based defensive poisoning against system identification attacks on cyber-physical systems

    PaperarXiv: 其他 AI 安全论文

  24. GateWatch: Low-overhead LLM monitoring with Boolean circuits

    PaperarXiv: 其他 AI 安全论文

  25. SCR: Detecting Backdoor-Poisoned Samples via Clean Reference Residuals

    PaperarXiv: 其他 AI 安全论文

  26. Probing agent over-proactiveness in real-world settings with Continuation Audits

    PaperarXiv: 其他 AI 安全论文

  27. Nothing to Declare: Steering a Single Neuron Increases Concealment in Agent Teams

    PaperarXiv: 其他 AI 安全论文

  28. PIPETTE: prompt injection purple teaming for robust LLM agents

    PaperarXiv: 其他 AI 安全论文

  29. Can Weight-Space Programming Change a Model's Learned Ethical Priorities?

    PaperarXiv: 其他 AI 安全论文

  30. AutoPI: Taxonomy-Guided Adaptive Red Teaming for Prompt Injection in LLM Agents

    PaperarXiv: 其他 AI 安全论文

  31. When shuffling fails: recovering tokens from obfuscated KV caches

    PaperarXiv: 其他 AI 安全论文

  32. What survives in the shift from text to pixels? Conditional protection against prompt injection in documents rendered as images

    PaperarXiv: 其他 AI 安全论文

  33. Learning to Coordinate: LLM-Guided Workflow Evolution for Token-Efficient Red Teaming

    PaperarXiv: 其他 AI 安全论文

  34. Guarding the kingdom of skilled agents in a computational AI dynasty: control-flow integrity for agent skills

    PaperarXiv: 其他 AI 安全论文

  35. VESPA: Verifiable Symbolic Policy Adaptation for Safe Reinforcement Learning

    PaperarXiv: 其他 AI 安全论文

  36. Intent contracts: taking tools out of the agent's hands

    PaperarXiv: 其他 AI 安全论文

  37. SCAFFOLD: Toward Credential-Preserving Evidence Transfer for Reusable LLM Agent Skills

    PaperarXiv: 其他 AI 安全论文

  38. When can decision models help LLM-driven network discovery?

    PaperarXiv: 其他 AI 安全论文

  39. Rank-one bypass of fixed linear erasure

    PaperarXiv: 其他 AI 安全论文

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

    PaperarXiv: 其他 AI 安全论文