TELUS Digital Releases 2026 GenAI Safety Benchmark Covering 34 Leading AI Models
Key Highlights
- The 2026 GenAI Safety Model Benchmark analyzed more than 620,000 adversarial tests across 34 leading AI models from 10 global providers.
- Vulnerability rates ranged from 1.3% to 93%, demonstrating that even advanced models can be manipulated into unsafe behavior under targeted attacks.
- The benchmark found that some open-source models outperformed proprietary alternatives, challenging assumptions about open-model safety risks.
TELUS Digital released its 2026 GenAI Safety Model Benchmark, the company’s largest AI security study to date, based on more than 620,000 adversarial tests across 34 leading AI models. The findings underscore a critical challenge for enterprises deploying generative AI: even advanced models can be manipulated into unsafe behavior under targeted attacks.
The benchmark evaluated models from 10 providers across North America, Europe and China, including Anthropic, OpenAI, Google, Meta, Alibaba, Baidu and Mistral. Vulnerability rates ranged from 1.3% to 93%, with no model proving completely immune to adversarial exploitation.
TELUS Digital found that larger and reasoning-based models generally demonstrated stronger safety performance, while smaller models were more susceptible to manipulation. The research also showed that open-source models are not inherently less secure than proprietary alternatives, and that geography is not a reliable predictor of model safety.
Among the study’s key findings, privacy, fraud and cybersecurity risks emerged as the most common vulnerabilities across models. Researchers also identified a recurring “refuse-but-engage” pattern, where models initially declined harmful requests but still provided information that could be misused.
The company says the findings reinforce the need for continuous AI security testing, layered safeguards and human oversight rather than relying solely on model providers’ built-in protections. TELUS Digital conducted the benchmark using its Fuel iX Fortify platform, which automates adversarial testing and maps vulnerabilities against frameworks such as OWASP, NIST AI RMF and MITRE ATLAS.
The full 2026 GenAI Safety Model Benchmark report is available at:
TELUS Digital GenAI Safety Benchmark 2026.
Source: TELUS Digital
Stay Connected with ISE Magazine
Subscribe to our newsletters and magazine for the latest telecom insights, explore the current issue for in-depth features and strategies, and register for upcoming webinars to learn directly from industry leaders.