AI Safety Mechanisms Are Largely Superficial, Experts and Users Warn

A detailed technical analysis argues that the safety guardrails built into major AI large language models are far easier to bypass than vendors publicly claim. Simple techniques such as rephrasing questions, using role-play prompts, switching languages, or splitting sensitive queries into smaller parts can consistently circumvent content filters in leading models. Beyond individual model weaknesses, the existence of hundreds of globally available AI models with vastly different — or nonexistent — safety standards means users can simply switch platforms when one refuses a request. Open-source models such as Meta's Llama, Mistral, and DeepSeek can be downloaded and run locally in minutes, entirely removing server-side moderation layers. The analysis concludes that AI safety is fundamentally a global governance challenge rather than a technical one, and that without internationally coordinated standards, any single company's or country's safety efforts remain isolated and easily sidestepped.