mlsanityData to 7 Oct 2026

LLM safety and security

12 directions · 6,765 papers · 53 this week

Busiest this week

Most new since Mon 5 Oct

Safety alignment, refusal and guardrails

1,071 papers14 this weekHigh confidence in this grouping: the papers form one clear topicHigh confidence in this grouping: the papers form one clear topic · About

  1. How Narrative Wrapping Affects LLM Refusal: A Cross-Language Benchmark and Defense
  2. Safe at One Loop, Risky at Another: Aligning Safety Across Recurrent Depths in Looped Language Models
  3. How to train your model organism
  4. SafeEvo: Deciphering the Safety Alignment Mechanism and Evolution in Language Models
  5. Quad-State Safety Evaluation of Open-Weight Large Language Models on Non-Canonical Inputs
Open →

All directions

Largest first · week from Mon 5 Oct

  1. Safety alignment, refusal and guardrailssafety, alignment, llms, harmfulHigh confidence in this grouping: the papers form one clear topic · About1,07114High confidence in this grouping: the papers form one clear topic
  2. Machine unlearningunlearning, machine, machine unlearning, knowledgeHigh confidence in this grouping: the papers form one clear topic · About1,0436High confidence in this grouping: the papers form one clear topic
  3. Prompt injection and agent securityprompt, agents, injection, prompt injectionMedium confidence in this grouping: mostly one topic, with blurry edges · About95911Medium confidence in this grouping: mostly one topic, with blurry edges
  4. Jailbreakingjailbreak, llms, attacks, attackHigh confidence in this grouping: the papers form one clear topic · About9115High confidence in this grouping: the papers form one clear topic
  5. Watermarking generated content and modelswatermarking, watermark, watermarks, textHigh confidence in this grouping: the papers form one clear topic · About7176High confidence in this grouping: the papers form one clear topic
  6. Backdoor and data poisoning attacksbackdoor, attacks, backdoor attacks, poisoningHigh confidence in this grouping: the papers form one clear topic · About7094High confidence in this grouping: the papers form one clear topic
  7. Membership inference attacksmembership, inference, membership inference, privacyHigh confidence in this grouping: the papers form one clear topic · About5262High confidence in this grouping: the papers form one clear topic
  8. Vision-language and text-to-image safetysafety, visual, multimodal, adversarialLow confidence in this grouping: the papers may belong to more than one topic, or the group may change · About5072Low confidence in this grouping: the papers may belong to more than one topic, or the group may change
  9. Federated unlearning and poisoningfederated, federated learning, unlearning, clientsHigh confidence in this grouping: the papers form one clear topic · About2130High confidence in this grouping: the papers form one clear topic
  10. Audio jailbreaks and speech backdoorsaudio, speech, attack, attacksHigh confidence in this grouping: the papers form one clear topic · About561High confidence in this grouping: the papers form one clear topic
  11. Graph unlearninggraph, graph unlearning, unlearning, nodeHigh confidence in this grouping: the papers form one clear topic · About330High confidence in this grouping: the papers form one clear topic
  12. Mixture-of-experts routing safetyrouting, moe, experts, expertHigh confidence in this grouping: the papers form one clear topic · About201High confidence in this grouping: the papers form one clear topic
1 new this week isn’t in a direction yet
  1. BiasFlow: Geometric Monitoring and Backbone Regularization for Spurious Feature Reliance

Weekly activity

Papers per week, 22 Jun – 5 Oct (last week partial)

  1. DirectionConfidencepapers16 wks
  2. Safety alignment, refusal and guardrailsHigh confidence in this grouping: the papers form one clear topic · AboutHigh confidence in this grouping: the papers form one clear topicPer week: 22 Jun: 18; 29 Jun: 10; 6 Jul: 8; 13 Jul: 5; 20 Jul: 15; 27 Jul: 11; 3 Aug: 14; 10 Aug: 17; 17 Aug: 14; 24 Aug: 20; 31 Aug: 15; 7 Sep: 10; 14 Sep: 15; 21 Sep: 7; 28 Sep: 23; 5 Oct: 141,071216
  3. Prompt injection and agent securityMedium confidence in this grouping: mostly one topic, with blurry edges · AboutMedium confidence in this grouping: mostly one topic, with blurry edgesPer week: 22 Jun: 18; 29 Jun: 9; 6 Jul: 8; 13 Jul: 12; 20 Jul: 11; 27 Jul: 12; 3 Aug: 19; 10 Aug: 10; 17 Aug: 8; 24 Aug: 15; 31 Aug: 18; 7 Sep: 13; 14 Sep: 13; 21 Sep: 13; 28 Sep: 21; 5 Oct: 11959211
  4. Machine unlearningHigh confidence in this grouping: the papers form one clear topic · AboutHigh confidence in this grouping: the papers form one clear topicPer week: 22 Jun: 6; 29 Jun: 10; 6 Jul: 7; 13 Jul: 2; 20 Jul: 6; 27 Jul: 9; 3 Aug: 8; 10 Aug: 9; 17 Aug: 11; 24 Aug: 11; 31 Aug: 10; 7 Sep: 7; 14 Sep: 8; 21 Sep: 11; 28 Sep: 27; 5 Oct: 61,043148
  5. Watermarking generated content and modelsHigh confidence in this grouping: the papers form one clear topic · AboutHigh confidence in this grouping: the papers form one clear topicPer week: 22 Jun: 4; 29 Jun: 5; 6 Jul: 7; 13 Jul: 5; 20 Jul: 3; 27 Jul: 4; 3 Aug: 7; 10 Aug: 4; 17 Aug: 5; 24 Aug: 6; 31 Aug: 5; 7 Sep: 5; 14 Sep: 5; 21 Sep: 5; 28 Sep: 11; 5 Oct: 671787
  6. JailbreakingHigh confidence in this grouping: the papers form one clear topic · AboutHigh confidence in this grouping: the papers form one clear topicPer week: 22 Jun: 6; 29 Jun: 2; 6 Jul: 3; 13 Jul: 3; 20 Jul: 5; 27 Jul: 4; 3 Aug: 2; 10 Aug: 5; 17 Aug: 3; 24 Aug: 5; 31 Aug: 9; 7 Sep: 7; 14 Sep: 3; 21 Sep: 5; 28 Sep: 11; 5 Oct: 591178
  7. Backdoor and data poisoning attacksHigh confidence in this grouping: the papers form one clear topic · AboutHigh confidence in this grouping: the papers form one clear topicPer week: 22 Jun: 9; 29 Jun: 4; 6 Jul: 3; 13 Jul: 0; 20 Jul: 3; 27 Jul: 3; 3 Aug: 6; 10 Aug: 5; 17 Aug: 2; 24 Aug: 5; 31 Aug: 3; 7 Sep: 5; 14 Sep: 1; 21 Sep: 5; 28 Sep: 11; 5 Oct: 470969
  8. Vision-language and text-to-image safetyLow confidence in this grouping: the papers may belong to more than one topic, or the group may change · AboutLow confidence in this grouping: the papers may belong to more than one topic, or the group may changePer week: 22 Jun: 6; 29 Jun: 5; 6 Jul: 5; 13 Jul: 2; 20 Jul: 4; 27 Jul: 9; 3 Aug: 2; 10 Aug: 4; 17 Aug: 9; 24 Aug: 4; 31 Aug: 8; 7 Sep: 3; 14 Sep: 0; 21 Sep: 4; 28 Sep: 9; 5 Oct: 250776
  9. Membership inference attacksHigh confidence in this grouping: the papers form one clear topic · AboutHigh confidence in this grouping: the papers form one clear topicPer week: 22 Jun: 3; 29 Jun: 4; 6 Jul: 0; 13 Jul: 2; 20 Jul: 0; 27 Jul: 3; 3 Aug: 1; 10 Aug: 0; 17 Aug: 3; 24 Aug: 2; 31 Aug: 2; 7 Sep: 8; 14 Sep: 0; 21 Sep: 1; 28 Sep: 4; 5 Oct: 252635
  10. Audio jailbreaks and speech backdoorsHigh confidence in this grouping: the papers form one clear topic · AboutHigh confidence in this grouping: the papers form one clear topicPer week: 22 Jun: 0; 29 Jun: 2; 6 Jul: 0; 13 Jul: 2; 20 Jul: 0; 27 Jul: 2; 3 Aug: 0; 10 Aug: 1; 17 Aug: 0; 24 Aug: 0; 31 Aug: 1; 7 Sep: 1; 14 Sep: 0; 21 Sep: 2; 28 Sep: 1; 5 Oct: 15613
  11. Mixture-of-experts routing safetyHigh confidence in this grouping: the papers form one clear topic · AboutHigh confidence in this grouping: the papers form one clear topicPer week: 22 Jun: 0; 29 Jun: 0; 6 Jul: 0; 13 Jul: 0; 20 Jul: 0; 27 Jul: 0; 3 Aug: 0; 10 Aug: 0; 17 Aug: 0; 24 Aug: 0; 31 Aug: 1; 7 Sep: 0; 14 Sep: 0; 21 Sep: 0; 28 Sep: 0; 5 Oct: 1202
  12. Federated unlearning and poisoningHigh confidence in this grouping: the papers form one clear topic · AboutHigh confidence in this grouping: the papers form one clear topicPer week: 22 Jun: 3; 29 Jun: 1; 6 Jul: 0; 13 Jul: 0; 20 Jul: 0; 27 Jul: 2; 3 Aug: 1; 10 Aug: 1; 17 Aug: 0; 24 Aug: 2; 31 Aug: 1; 7 Sep: 0; 14 Sep: 0; 21 Sep: 0; 28 Sep: 0; 5 Oct: 021311
  13. Graph unlearningHigh confidence in this grouping: the papers form one clear topic · AboutHigh confidence in this grouping: the papers form one clear topicPer week: 22 Jun: 0; 29 Jun: 0; 6 Jul: 0; 13 Jul: 0; 20 Jul: 0; 27 Jul: 1; 3 Aug: 0; 10 Aug: 0; 17 Aug: 0; 24 Aug: 2; 31 Aug: 0; 7 Sep: 1; 14 Sep: 0; 21 Sep: 0; 28 Sep: 0; 5 Oct: 0334