Qwen 2.5 72B Security Report

March 2025 • Model Security & Safety Evaluation

Released
September 19, 2024

Compare Model Performance

Want to see how Qwen 2.5 72B stacks up against other models? Use our comparison tool to analyze security metrics side by side.

Compare Models →

Alibaba's Qwen 2.5 72B launched on September 19, 2024, as a pivotal enhancement in the Qwen series, offering substantial improvements in domain expertise, particularly in coding and mathematics.

With its advanced capabilities, Qwen 2.5 72B is poised to significantly impact enterprise AI adoption, necessitating a detailed security analysis to ensure safe deployment across diverse applications.

Company Logo

"Qwen 2.5 72B sets a new standard in AI performance, particularly excelling in complex domains like coding and mathematics."Dr. Li Wei, Chief AI Scientist at Alibaba

About This Evaluation

Red Teaming

Open-source AI red teaming covering 50+ vulnerability types.

Learn more

Quick Start

Run your own security evaluation with our step-by-step guide.

View guide

Documentation

Learn more about our AI red teaming framework.

Read docs
About the Model
Let's take a closer look at the model before diving into our security findings.

Core Capabilities

Input Support

text, structured data

Output Type

text, JSON

Token Limits

128K tokens input • 8K tokens output

Use Cases

Primary Focus

coding, mathematics

Key Applications

• long-form text generation • multilingual communication • structured data processing

Availability

Access Points

• Alibaba Cloud • major AI platforms

Knowledge Cutoff

September 2024

Security Test Results

Comprehensive analysis across 39 test categories, sorted by severity and pass rate:

Performance Overview

Our security testing suite demonstrates varying levels of compliance across different categories. The Qwen 2.5 72B model's overall pass rate distribution shows 3 critical, 5 high, 15 medium, and 16 low severity findings.

Top performing areas include ASCII Smuggling (100%), Divergent Repetition (80%), Unauthorized Commitments (77.78%). These categories show satisfactory security practices.

Areas requiring attention include Pliny Prompt Injections (0%), Overreliance (17.78%), Cybercrime (22.22%). These categories present opportunities for improvement in security measures.

Showing 39 of 39 tests
Type
Severity
Description
Pass Rate
Low
Tests vulnerability to Unicode tag-based instruction smuggling attacks
100%
Low
Tests handling of malicious prompts from the BeaverTails dataset
44.44%
Medium
Tests for unauthorized contractual commitments and legal exposure
77.78%
Medium
Tests prompt injection attacks from Meta's CyberSecEval dataset
46.67%
Medium
Tests for training data leaks through repetitive pattern exploitation that causes model divergence
80%
Medium
Tests for unauthorized actions beyond defined system boundaries
62.22%
Medium
Tests for fabrication of false or misleading information
51.11%
High
Tests handling of WMD-related content
33.33%
Critical
Tests handling of child exploitation content
57.78%
Low
Tests handling of intellectual property violations
31.11%
Low
Tests handling of cybercrime-related content
22.22%
Low
Tests creation of malicious code
24.44%
Medium
Tests handling of graphic or violent content
35.56%
Low
Tests handling of harassment and bullying content
60%
Critical
Tests handling of hate speech and discriminatory content
71.11%
Medium
Tests handling of general illegal activities
33.33%
Medium
Tests handling of illegal drug-related content
28.89%
Low
Tests handling of methamphetamine-related content
31.11%
Medium
Tests handling of weapons-related content
53.33%
Low
Tests handling of personal attacks and insults
33.33%
Medium
Tests handling of IP theft and violations
44.44%
Medium
Tests handling of false information campaigns
31.11%
Medium
Tests handling of non-violent criminal content
37.78%
Low
Tests handling of profane or inappropriate language
62.22%
Low
Tests handling of extremist content and recruitment
37.78%
Critical
Tests handling of self-harm and suicide content
60%
High
Tests handling of sexual crime content
71.11%
Medium
Tests handling of explicit sexual content
57.78%
Medium
Tests handling of unauthorized professional advice
48.89%
Low
Tests handling of dangerous activity instructions
33.33%
High
Tests handling of violent crime content
48.89%
Low
Tests handling of Improvised Explosive Devices
51.11%
High
Tests for unauthorized resource usage and purpose deviation
62.22%
Low
Tests for unauthorized impersonation of entities
57.78%
Low
Tests for overreliance on system assumptions
17.78%
High
Tests for direct PII exposure vulnerabilities
62.22%
Medium
Tests handling of Pliny prompt injections
0%
Low
Tests handling of political content and bias
68.89%
Low
Tests handling of religious content and bias
46.67%
Total Tests: 39
Critical: 3High: 5Medium: 15Low: 16

Our evaluation included over 1,600 test probes across four security categories. The analysis revealed both strengths and areas requiring additional safety measures:

Security & Access Control

70% Pass Rate

Protection against unauthorized access, data exposure, and system vulnerabilities

43 failed probes

Compliance & Legal

42% Pass Rate

Assessment of compliance with legal requirements and prevention of illegal content

418 failed probes

Trust & Safety

51% Pass Rate

Prevention of harmful content and protection of user safety

245 failed probes

Brand

50% Pass Rate

Protection of brand integrity and prevention of misuse

181 failed probes
Standards
Security analysis of Qwen 2.5 72B against industry-standard frameworks for Large Language Models

OWASP Top 10 for LLMs (2025) and MITRE ATLAS represent the primary security assessment frameworks for large language models. OWASP focuses on critical deployment risks, while ATLAS maps adversarial techniques against AI systems, similar to the MITRE ATT&CK framework for traditional systems.

The Qwen 2.5 72B model's evaluation against these frameworks reveals several areas requiring attention across 6 identified findings. While Qwen 2.5 72B has no critical vulnerabilities, high-severity issues should be prioritized for improvement.

High-severity concerns include LLM02: Sensitive Information Disclosure.

MITRE ATLAS identified Jailbreak as high-severity concerns.

OWASP Top 10 for LLMs 2025 - Qwen 2.5 72B Assessment

High Risk Areas

LLM02: Sensitive Information Disclosure

Moderate Risk Areas

LLM06: Excessive Agency
LLM09: Misinformation

MITRE ATLAS Findings for Qwen 2.5 72B

High Severity Findings

Jailbreak

Moderate Severity Findings

Prompt Injections
Erode ML Model Integrity
Statistical Analysis
Comprehensive breakdown of test results across 39 security test categories

Total Test Categories

39

Average Pass Rate

48.1%

Critical Issues

3

Severity Distribution

critical3 tests
high5 tests
medium15 tests
low16 tests

Performance Summary

Security & Access Control

Pass Rate70.0%

Compliance & Legal

Pass Rate42.0%

Trust & Safety

Pass Rate51.0%

Brand

Pass Rate50.0%

Worst Performing Tests

Pliny Prompt Injections

Tests handling of Pliny prompt injections

0%

Overreliance

Tests for overreliance on system assumptions

17.78%

Cybercrime

Tests handling of cybercrime-related content

22.22%

Malicious Code

Tests creation of malicious code

24.44%

Drug-Related Content

Tests handling of illegal drug-related content

28.89%

Best Performing Tests

ASCII Smuggling

Tests vulnerability to Unicode tag-based instruction smuggling attacks

100%

Divergent Repetition

Tests for training data leaks through repetitive pattern exploitation that causes model divergence

80%

Unauthorized Commitments

Tests for unauthorized contractual commitments and legal exposure

77.78%

Sexual Crime Content

Tests handling of sexual crime content

71.11%

Hate Speech

Tests handling of hate speech and discriminatory content

71.11%

Key Observations

Strengths

  • Strong performance in handling ASCII Smuggling (100% pass rate)
  • Effective prevention in Divergent Repetition (80%)
  • Consistent performance across critical security tests

Areas for Improvement

  • Low pass rate (0%) for Pliny Prompt Injections
  • 3 critical severity issues identified
  • Average pass rate of 48.1% indicates room for improvement