Cisco Reports Vulnerabilities in Open-Weight LLMs to Multi-Turn Attacks

First seen 10 Nov 2025, 11:52 UTC • Infosecurity-Magazine

•71% similarity •17

Ask AI about this cluster

Article Content

Include sub-articles 1 / 1

Browse articles

ThreatCluster

Cisco's research reveals that popular open-weight large language models (LLMs) are highly vulnerable to multi-turn prompt injection attacks. These adaptive adversarial attacks can achieve success rates exceeding 90% against most tested defenses, despite single-turn defenses appearing robust. The findings highlight significant weaknesses in the security of widely used generative AI services.

ThreatCluster AI

Attack Flow

10 entities, 12 inferred relationships (STIX 2.1).

Technique (2)

Data Breach
Prompt Injection

Affected Platform (8)

Alibaba Qwen3-32B
DeepSeek v3.1
Google Gemma-3-1B-1T
Meta Llama 3.3-70B-Instruct
Microsoft Phi-4
Mistral Large-2
OpenAI GPT-OSS-2-B
Zhipu AI GLM-4.5-Air

Relationships

Data Breach targets Alibaba Qwen3-32B
Data Breach targets DeepSeek v3.1
Data Breach targets Google Gemma-3-1B-1T
Data Breach targets Meta Llama 3.3-70B-Instruct
Prompt Injection targets Alibaba Qwen3-32B
Prompt Injection targets DeepSeek v3.1
Prompt Injection targets Google Gemma-3-1B-1T
Prompt Injection targets Meta Llama 3.3-70B-Instruct
Data Breach related to Microsoft Phi-4
Data Breach related to Mistral Large-2
Data Breach related to OpenAI GPT-OSS-2-B
Data Breach related to Zhipu AI GLM-4.5-Air

Community

Browse all →

Tracked Entities in This Story

Alibaba Qwen3-32B DeepSeek v3.1 Google Gemma-3-1B-1T Meta Llama 3.3-70B-Instruct Microsoft Phi-4 Mistral Large-2 Data Breach Prompt Injection

Cisco Reports Vulnerabilities in Open-Weight LLMs to Multi-Turn Attacks

Ask AI about this cluster

Article Content

More articles in this cluster

Attack Flow

Defensive Countermeasures

Extracted Entities

Community

Continue Reading

Tracked Entities in This Story