✨GateBreaker: Gate-Guided Attacks on Mixture-of-Expert LLMs
📝 Summary:
GateBreaker is the first framework to compromise MoE LLM safety by identifying and disabling ~3% of safety neurons in expert layers. This raises attack success rates from 7.4% to 64.9% across eight LLMs and generalizes to VLMs, showing concentrated and transferable safety vulnerabilities.
🔹 Publication Date: Published on Dec 24
🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.21008
• PDF: https://arxiv.org/pdf/2512.21008
==================================
For more data science resources:
✓ https://xn--r1a.website/DataScienceT
#LLM #AIsecurity #MoELLMs #AIvulnerability #GateBreaker
📝 Summary:
GateBreaker is the first framework to compromise MoE LLM safety by identifying and disabling ~3% of safety neurons in expert layers. This raises attack success rates from 7.4% to 64.9% across eight LLMs and generalizes to VLMs, showing concentrated and transferable safety vulnerabilities.
🔹 Publication Date: Published on Dec 24
🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.21008
• PDF: https://arxiv.org/pdf/2512.21008
==================================
For more data science resources:
✓ https://xn--r1a.website/DataScienceT
#LLM #AIsecurity #MoELLMs #AIvulnerability #GateBreaker
✨Exposing the Systematic Vulnerability of Open-Weight Models to Prefill Attacks
📝 Summary:
A study reveals prefill attacks as a critical, underexplored vulnerability in open-weight language models. These attacks, which predefine initial response tokens, consistently compromise major models, necessitating urgent defense development.
🔹 Publication Date: Published on Feb 16
🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.14689
• PDF: https://arxiv.org/pdf/2602.14689
==================================
For more data science resources:
✓ https://xn--r1a.website/DataScienceT
#PrefillAttacks #LLMSecurity #AIvulnerability #OpenWeightModels #LanguageModels
📝 Summary:
A study reveals prefill attacks as a critical, underexplored vulnerability in open-weight language models. These attacks, which predefine initial response tokens, consistently compromise major models, necessitating urgent defense development.
🔹 Publication Date: Published on Feb 16
🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.14689
• PDF: https://arxiv.org/pdf/2602.14689
==================================
For more data science resources:
✓ https://xn--r1a.website/DataScienceT
#PrefillAttacks #LLMSecurity #AIvulnerability #OpenWeightModels #LanguageModels
✨AgentHazard: A Benchmark for Evaluating Harmful Behavior in Computer-Use Agents
📝 Summary:
Computer-use agents pose unique safety risks as harm can emerge from sequences of individually benign actions. AgentHazard is a benchmark with 2,653 instances to evaluate this. Experiments reveal current systems are highly vulnerable, showing model alignment alone doesnt ensure agent safety.
🔹 Publication Date: Published on Apr 3
🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.02947
• PDF: https://arxiv.org/pdf/2604.02947
• Project Page: https://yunhao-feng.github.io/AgentHazard/
==================================
For more data science resources:
✓ https://xn--r1a.website/DataScienceT
#AISafety #AgentAI #AIVulnerability #AIethics #AIbenchmark
📝 Summary:
Computer-use agents pose unique safety risks as harm can emerge from sequences of individually benign actions. AgentHazard is a benchmark with 2,653 instances to evaluate this. Experiments reveal current systems are highly vulnerable, showing model alignment alone doesnt ensure agent safety.
🔹 Publication Date: Published on Apr 3
🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.02947
• PDF: https://arxiv.org/pdf/2604.02947
• Project Page: https://yunhao-feng.github.io/AgentHazard/
==================================
For more data science resources:
✓ https://xn--r1a.website/DataScienceT
#AISafety #AgentAI #AIVulnerability #AIethics #AIbenchmark