✨Privacy Collapse: Benign Fine-Tuning Can Break Contextual Privacy in Language Models

📝 Summary:
Benign fine-tuning can cause privacy collapse in language models. Models lose contextual privacy reasoning despite maintaining high performance, leading to severe vulnerabilities. This silent failure reveals a critical gap in current safety evaluations for specialized agents.

🔹 Publication Date: Published on Jan 21

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.15220
• PDF: https://arxiv.org/pdf/2601.15220
• Github: https://github.com/parameterlab/privacy-collapse

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#LLM #Privacy #AIsafety #FineTuning #AIsecurity

✨Less Is More -- Until It Breaks: Security Pitfalls of Vision Token Compression in Large Vision-Language Models

📝 Summary:
Visual token compression degrades LVLM robustness via unstable token importance ranking. This causes critical information loss, creating vulnerabilities only under compression. An attack exploits this, revealing an efficiency-security trade-off.

🔹 Publication Date: Published on Jan 17

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.12042
• PDF: https://arxiv.org/pdf/2601.12042

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#LVLM #AIsecurity #VisionAI #ModelRobustness #DeepLearning

✨Spider-Sense: Intrinsic Risk Sensing for Efficient Agent Defense with Hierarchical Adaptive Screening

📝 Summary:
Spider-Sense is an event-driven framework for agent security using Intrinsic Risk Sensing. It provides intrinsic, selective defense through a hierarchical mechanism, activating only upon risk perception. It achieves low attack success and false positive rates with minimal latency.

🔹 Publication Date: Published on Feb 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.05386
• PDF: https://arxiv.org/pdf/2602.05386
• Github: https://github.com/aifinlab/Spider-Sense

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#Cybersecurity #AgentSecurity #AISecurity #RiskSensing #AutonomousAgents

✨GoodVibe: Security-by-Vibe for LLM-Based Code Generation

📝 Summary:
GoodVibe secures LLM-generated code by precisely fine-tuning only a small subset of security-relevant neurons. This neuron-level framework greatly enhances code security and preserves utility with significantly fewer parameters and training costs than traditional methods.

🔹 Publication Date: Published on Feb 11

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.10778
• PDF: https://arxiv.org/pdf/2602.10778

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#LLM #CodeGeneration #Cybersecurity #AIsecurity #MachineLearning

✨Visual Memory Injection Attacks for Multi-Turn Conversations

📝 Summary:
Visual Memory Injection VMI covertly manipulates generative vision-language models using images. These images trigger specific manipulative responses only with certain prompts in multi-turn conversations, showing large-scale user manipulation is feasible.

🔹 Publication Date: Published on Feb 17

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.15927
• PDF: https://arxiv.org/pdf/2602.15927
• Github: https://github.com/chs20/visual-memory-injection

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#VMI #VisionLanguageModels #AISecurity #AIManipulation #GenerativeAI

✨ProtegoFed: Backdoor-Free Federated Instruction Tuning with Interspersed Poisoned Data

📝 Summary:
ProtegoFed is a new federated instruction tuning framework. It detects and removes widespread poisoned data across clients using frequency domain gradient analysis and collaborative clustering, reducing attack success to almost zero while maintaining utility.

🔹 Publication Date: Published on Feb 28

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.00516
• PDF: https://arxiv.org/pdf/2603.00516
• Project Page: https://github.com/dongdongzhaoUP/ProtegoFed
• Github: https://github.com/dongdongzhaoUP/ProtegoFed

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#FederatedLearning #AIsecurity #DataPoisoning #MachineLearning #AIResearch

✨SlowBA: An efficiency backdoor attack towards VLM-based GUI agents

📝 Summary:
SlowBA is a novel backdoor attack targeting the response latency of VLM-based GUI agents. It induces excessively long reasoning chains using realistic pop-up window triggers, significantly increasing response length and latency while maintaining task accuracy. This reveals a new security vulnerab...

🔹 Publication Date: Published on Mar 9

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.08316
• PDF: https://arxiv.org/pdf/2603.08316
• Github: https://github.com/tu-tuing/SlowBA

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#BackdoorAttack #AISecurity #VLM #GUIagents #Cybersecurity

✨Your Agent, Their Asset: A Real-World Safety Analysis of OpenClaw

📝 Summary:
A real-world safety analysis of the personal AI agent OpenClaw reveals significant vulnerabilities due to its broad system access. Attacks targeting its Capability, Identity, or Knowledge CIK dimensions drastically increase success rates, and current defenses are insufficient, indicating inherent...

🔹 Publication Date: Published on Apr 6

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.04759
• PDF: https://arxiv.org/pdf/2604.04759
• Project Page: https://ucsc-vlaa.github.io/CIK-Bench/
• Github: https://github.com/UCSC-VLAA/CIK-Bench

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#AISafety #Cybersecurity #AIAgents #Vulnerability #AIsecurity

✨Backdoor Attacks on Decentralised Post-Training

📝 Summary:
This paper introduces the first backdoor attack on pipeline parallelism in decentralized LLM post-training. An adversary controlling an intermediate stage can significantly misalign the model, reducing alignment from 80% to 6% with a trigger word, even resisting safety training.

🔹 Publication Date: Published on Mar 31

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.02372
• PDF: https://arxiv.org/pdf/2604.02372

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#BackdoorAttack #LLM #DecentralizedAI #AISecurity #MachineLearning

✨Turing Test on Screen: A Benchmark for Mobile GUI Agent Humanization

📝 Summary:
This paper introduces the Turing Test on Screen to address GUI agents detectability by digital platforms. It proposes a benchmark and methods to humanize agent behavior, balancing imitability with task performance, enabling seamless coexistence in adversarial digital environments.

🔹 Publication Date: Published on Feb 24

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.09574
• PDF: https://arxiv.org/pdf/2604.09574

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#TuringTest #GUIAgents #AIHumanization #MobileAI #AISecurity