#USA #Vancouver #AI Safety #TELUS Digital #GenAI Benchmark

TELUS Digital Unveils Comprehensive AI Safety Findings

In a groundbreaking study, TELUS Digital recently released its GenAI Safety Model Benchmark, revealing significant insights into the vulnerabilities associated with generative AI (GenAI) applications. Conducted with over 620,000 adversarial tests across 34 different AI models from notable global providers, this research serves as an essential guide for enterprises navigating the complexities of AI safety.

The reality is that AI models, while powerful, are not immune to adversarial attacks. The study illustrates how certain models can fall prey to harmful requests over 90% of the time. Hence, enterprises must understand that the deployment of AI carries inherent risks, which can be mitigated through rigorous testing and proactive security measures.

The Scale of the Benchmark

This study, a follow-up to TELUS Digital's initial benchmark published in 2025, marks an expansion in scope, analyzing models from key players spanning North America, Europe, and China. Notable components include models such as Claude from Anthropic, GPT from OpenAI, and LLaMA from Meta. The findings indicate that while newer models tend to exhibit improved security features, the quality of safety is not guaranteed and depends on various factors, including the design approach of the model and its size.

Interestingly, smaller models consistently demonstrated lower resilience against attacks, reinforcing the notion that size alone is not a distinguishing factor for security. More sophisticated reasoning models outperformed simpler versions, indicating that how an AI functions internally plays a vital role in determining its vulnerability.

Key Findings and Insights

Bret Kinsella, TELUS Digital's General Manager and Senior Vice President, emphasized the need for organizations to transition from mere sporadic checks of AI security at the time of deployment to a more dynamic and ongoing testing strategy. The research identified three critical predictors of AI safety:
1. Model Design: Models with advanced reasoning capabilities were harder to exploit, showing merged resilience against adversarial methods.
2. Model Size: Generally, smaller models (both open-source and proprietary) reflected greater vulnerabilities, pushing the need for cautious selection of AI applications.
3. Development Approach: The methodologies employed by developers significantly affect model safety and must be taken into account during the deployment phase.

Notably, testing processes further revealed that open-source models are not inherently more dangerous than their proprietary counterparts. While they experienced slightly more attacks on average, exceptions like Zhipu AI’s GLM model performed admirably in safety metrics, outperforming many paid versions.

Moving Forward: The Path to Enhanced AI Security

The implications of TELUS Digital's research highlight that AI vulnerabilities are not all equal; some pose a more severe risk to privacy and are exploited for fraudulent purposes. Organizations must not only hone in on the ongoing vulnerabilities identified in this report but also adopt recommended guardrails to safeguard their AI applications against misuse effectively.

Investing in AI security should be a top priority. Current estimates suggest global AI spending will total $2.52 trillion by 2026, but a mere $3.43 billion is allocated toward securing it. Such disparities underscore the need for immediate action in enhancing AI security measures.

As enterprises navigate this challenging landscape, the TELUS GenAI Safety Model Benchmark advocates for layered defense strategies that encompass comprehensive testing and human expertise. Effective AI application management, including prompt shielding and data governance, can greatly reduce exploit risks and strengthen overall safety performance.

TELUS Digital’s continuous security testing solution, Fuel iX Fortify, integrates automated testing with in-depth evaluations against industry standards, empowering organizations to proactively mitigate risks. By marrying technology with human oversight, TELUS Digital reinforces the foundation necessary for securing AI environments effectively.

For those looking to learn more about the GenAI Safety Model Benchmark's insights and its implications, the full report is accessible online. With the urgent need for robust AI security protocols, TELUS Digital stands at the forefront, paving the way for responsible and compliant enterprise AI deployment.