Cloudflare, a leading global internet security company, has expressed serious concerns about the inconsistent way many artificial intelligence models handle harmful or dangerous requests during security tests.

In a new report released on Thursday, Cloudflare revealed that leading AI systems often give conflicting responses when asked to perform potentially risky tasks. The same model would sometimes refuse a dangerous request but accept a similar one when phrased differently. This inconsistency, according to the company, makes it difficult to trust current AI safety measures.

Cloudflare’s security team tested several major AI models using various prompts designed to check their safety boundaries. The results showed that many models failed to consistently reject harmful instructions, raising questions about their reliability in real-world applications.

The company warned that such unpredictability could create serious security risks, especially for businesses and organisations that rely on AI for sensitive operations. Inconsistent refusals could be exploited by attackers to bypass safety guardrails and generate dangerous content or instructions.

“This inconsistency is more than a technical issue  it’s a security vulnerability,” a Cloudflare executive stated. “Organisations need to understand that today’s AI safety systems are still not dependable enough for high-stakes environments.”

 As AI adoption grows rapidly across Africa and the world, concerns about safety, ethics, and security have increased. Many African businesses and governments are beginning to integrate AI into critical sectors such as healthcare, finance, education, and cybersecurity. Inconsistent safety behaviour in these models could expose users to risks ranging from misinformation to more serious threats.

Cloudflare urged AI developers to strengthen their safety mechanisms and called for more transparent testing standards across the industry.

 The report serves as a timely reminder that while AI continues to advance, its safety systems still require significant improvement. As more African countries embrace artificial intelligence, experts say greater attention must be paid to building trustworthy and secure systems that can protect users and organisations.