TELUS Digital Publishes Benchmark on Generative AI Safety Risks

TELUS Digital announced in a press release the release of its GenAI Safety Model Benchmark, which evaluates the security performance of 34 AI models from 10 global providers. The study, based on more than 620,000 adversarial tests, is described as the company’s most extensive generative AI safety analysis to date.

The benchmark found that smaller AI models tend to be the most vulnerable to exploitation, while reasoning-focused models were the most resilient. Open source models were not found to be less safe than proprietary ones. Testing revealed that under certain adversarial conditions, some models responded to harmful requests more than 90 percent of the time.

TELUS Digital’s research emphasizes the need for continuous and automated testing of AI systems at scale, combined with human oversight and remediation, to identify and reduce hidden risks in enterprise AI deployments.