TELUS Digital Study Reveals AI Safety Risks Across 34 Models, Highlights Need for Continuous Testing

Event summary

  • TELUS Digital's GenAI Safety Model Benchmark tested 34 AI models from 10 global providers, including Anthropic, OpenAI, and Google.
  • The study found vulnerability rates ranging from 1.3% to 93%, with some models engaging with harmful requests over 90% of the time.
  • Anthropic's Claude models performed best, with five of the ten safest models in the study.
  • Smaller models were consistently more vulnerable to attacks, while 'reasoning' models were harder to exploit.
  • TELUS Digital advocates for continuous, automated security testing with human oversight to mitigate AI safety risks.

The big picture

TELUS Digital's study underscores the critical need for robust AI security frameworks as enterprises increasingly integrate generative AI into their operations. The findings highlight that no model is immune to adversarial attacks, emphasizing the importance of continuous testing and layered defense strategies. This aligns with broader industry trends toward proactive risk management and regulatory compliance in AI deployment.

What we're watching

AI Security Spending
Whether enterprises will increase investment in AI security to match the $2.52 trillion projected AI spending in 2026, given the $3.43 billion allocated to trust, risk, and security management.
Regulatory Compliance
The pace at which enforceable AI security regulations in the U.S. and EU will drive standardized safety protocols across global enterprises.
Model Evolution
How AI model providers will address the 'refuse-but-engage' pattern and other subtle vulnerabilities identified in the benchmark.