Insecure about GPT-5

OpenAI’s latest large language model (LLM), GPT-5, came out with much fanfare on Aug. 7, but less than a week later, most of the chat about ChatGPT’s newest engine has not been too kind, to say the least.

Billed as a more intelligent, reliable, and versatile AI system that can handle a wider range of complex tasks, with a focus on deep reasoning, multimodal understanding, and agentic capabilities, users have not taken too kindly to the new LLM, calling it “awful” and a “mess,” among the comments I’ve seen on Reddit over the past week.

So, what are users, techies and researchers in the AI community upset about regarding OpenAI’s latest creation? In addition to underwhelming performance, user experience and reliability problems, and ethical concerns, security vulnerabilities and weaknesses revealed the past several days have raised concerns over the LLM’s ability to protect against various risks.

In fact, cybersecurity researchers have quickly demonstrated vulnerabilities in GPT-5, identifying several methods to bypass its safety mechanisms and elicit harmful or undesired outputs. 

NeuralTrust – a security platform for generative AI – was able to jailbreak GPT-5 by using its "Echo Chamber and Storytelling” technique to prompt it to produce directions for building a Molotov cocktail.

Another AI security platform, SPLX, found significant security and safety weaknesses in GPT-5’s default configuration. SPLX’s testing with over 1,000 adversarial prompts showed that the default GPT-5 model scored poorly on security (2.4%), safety (13.6%), and business alignment (1.7%). 

This prompted SPLX to call GPT-5’s raw model “nearly unusable for enterprise out of the box” due to vulnerabilities to various attacks, including prompt injection, data exfiltration, and jailbreaking. 

In addition, SPLX warned, “OpenAI’s latest model is undeniably impressive, but security and alignment must still be engineered, not assumed.”

On the other side of the coin, Dr. Sarah Bird, chief product officer of responsible AI at Microsoft, raved about GPT-5’s security features. “GPT-5 is not only a huge step forward in innovation but has one of the strongest safety profiles of any Open AI model,” she boasted.

Well, Dr. Bird, it appears a vast majority of folks who used and tested GPT-5 over the past week beg to differ.

Even OpenAI CEO Sam Altman admitted it was a rough rollout for GPT-5, citing the breakdown of the autoswitcher between GPT-5 and its predecessor, GPT-4o. “The result was GPT-5 seemed way dumber,” he said, adding that “it was a little more bumpy than we hoped for!”

Altman promised to implement fixes to “get things stable” and improve GPT-5’s performance and the overall user experience.

I hope that security is near the top of that list of GPT-5 fixes. Based on the feedback and research I’ve seen, that may not necessarily be the case.