TELUS Digital Benchmark Finds AI Model Vulnerabilities Ranging From 1.3% to 93%
TU•TELUS Digital's GenAI Safety Model Benchmark tested 34 AI models from ten global providers with over 620,000 adversarial attacks, revealing vulnerability rates spanning 1.3% to 93%. The study found reasoning models least exploitable, smaller models most at risk and open-source models not inherently less secure than proprietary alternatives.
1. Benchmark Scope and Methodology
TELUS Digital conducted its second GenAI Safety Model Benchmark using more than 620,000 adversarial tests across 34 leading AI models from ten providers in North America, Europe and China. This edition nearly doubles the scope of the November 2025 study, expanding open-source tests from two to fourteen and covering models such as Claude, GPT, Gemini, LLaMA, Qwen and ERNIE.
2. Vulnerability Findings
Overall attack vulnerability rates ranged from a low of 1.3% to a high of 93%, with ten models scoring below 5% failure, including five of Anthropic’s Claude variants. The research highlighted that reasoning models averaged a 19.9% vulnerability rate, while smaller, budget-friendly models were most susceptible and open-source alternatives sometimes outperformed proprietary counterparts.
3. Model Safety Factors
Key predictors of safety included the model’s reasoning capabilities, size and development approach. Reasoning models, which process queries before answering, proved significantly harder to exploit than standard models. Larger models typically showed greater resistance, but high performance did not always correlate with improved safety.
4. Enterprise Implications
TELUS Digital recommends continuous, automated security testing with human oversight to uncover hidden vulnerabilities that spot checks miss. Enterprises are urged to implement guardrails, regular adversarial assessments and configuration reviews to maintain a robust security posture when deploying GenAI solutions.
