# AI safety benchmark reveals deeper LLM weaknesses > Source: > Published: 2026-05-29 06:01:30+00:00 Another day, another benchmark uncovers new risks and vulnerabilities in language models. Tech services firm TELUS Digital has [released the results](https://www.telusdigital.com/about/newsroom/telus-digital-research-uncovers-genai-risks-and-offers-blueprint-to-protect-enterprise-ai-applications) of one of the most comprehensive AI safety and cybersecurity benchmarks to date. The benchmark used 620,000 attack simulations and tested them on 34 models from 10 global AI labs. Models tested included: Claude, GPT, Gemini, LlaMA, Mistral, Qwen (Alibaba), ERNIE (Baidu), Seed (ByteDance), and GLM (Z.ai). "The findings highlight an important reality for enterprises deploying AI: with the right adversarial techniques, AI models can be coaxed into unsafe behavior," the report stated. It also found that "some models engaged with harmful requests more than 90% of the time." Some of the key findings included: **Reasoning models are safer**: The reasoning models that think through their responses before answering turned out to be far safer in this study. They were vulnerable to 19.9% of attacks, compared to non-reasoning models that were vulnerable to 55.1% of attacks.**SLMs are easier to crack**: Smaller, more economical models struggled in this benchmark as TELUS reported they "were far more likely to be exploited than their larger counterparts."**Open models aren't always less safe**: While the benchmark found open models fell victim to exploits more often than proprietary ones did, it also found that a large open-source model, GLM 4.7 from Z.ai, outperformed many proprietary models on safety.**Models still stumble on three big risk factors**: The benchmark discovered that many of the models made progress on risk areas such as political manipulation since the[last version of the benchmark in November 2025](https://www.telusdigital.com/insights/fuel-ix/resource/gen-ai-safety). Unfortunately, even the top performing models struggled in three areas: cybersecurity threats, privacy exploitation, and fraud. “The real risk isn't that AI models have vulnerabilities. It's that most organizations have no way of knowing which vulnerabilities apply to them,” said Bret Kinsella, Senior Vice President at TELUS Digital in the release. While the report uncovered new soft spots in models, that's also a key aspect of hardening any technology system: deeply understanding the risks so that you can fully prepare to mitigate them with multi-layered defense strategies. The report noted, "The encouraging news is that the research points to a clear path forward. The benchmark shows the importance of testing AI systems at scale to uncover hidden risks that may appear safe under less rigorous investigation. Continuous, automated security testing with human oversight and remediation can dramatically reduce risk." ## Our Deeper *View* If you're not convinced by now that you personally, and your company more broadly, need to put safeguards around your work with generative AI models and agents, then you should at least be conscious that you are taking on those risks and you're willing to do it because you've calculated that the potential trade-off are worth it. However, most people and organizations will find that better understanding the risks around language models will lead them to systematically mitigate those risks with a series of solutions, guardrails, and policies. AI is advancing so rapidly that adapting operations to manage risk is one of the most daunting tasks in the enterprise right now. But it's not slowing down or getting any easier, and so organizations are going to have to accept this as the new reality and find ways to get creative in order to keep up.