Tenable researchers have bypassed OpenAI’s new GPT-5 safety features within a day of its release. Using the “crescendo technique,” they prompted the model into giving instructions on making a Molotov cocktail by posing as a history student.
OpenAI launched GPT-5 on August 7, 2025, claiming stronger protections against harmful use. Tenable’s quick jailbreak, achieved in four prompts, shows the model still vulnerable to manipulation.
Tenable VP Tomer Avni warned this poses major security, ethical, and compliance risks, especially as employees adopt AI without oversight. He stressed that organisations need dedicated AI exposure management strategies instead of relying on built-in safeguards.
OpenAI says fixes are underway, but Tenable’s test underscores the need for external monitoring and governance of AI tools.
