OpenAI released GPT-5, its latest large language model, to the public on August 7. The model was touted as faster, smarter, and more capable than previous models, but security researchers quickly discovered its limitations. A recent study found that GPT-5 fails in core security and safety metrics, with vulnerabilities identified by outside security researchers that were already patched in older models.
A cybersecurity firm, SPLX, subjected GPT-5 to over 1,000 different attack scenarios, including prompt injection, data poisoning, jailbreaking, and data exfiltration. The results showed that the default version of GPT-5 was “nearly unusable for enterprises” out of the box, scoring only 2.4% in an assessment for security.
Other researchers have also identified significant vulnerabilities in GPT-5, including a way to jailbreak the base model through context poisoning. A cybersecurity firm, NeuralTrust, found that attackers can manipulate the contextual information and instructions used by GPT-5 to learn more about specific projects or tasks, allowing them to break free of its constraints.
The study’s authors concluded that “AI-driven automation comes with a profound security cost,” and demonstrated techniques to inject telemetry data into system management tools, leading to compromised integrity of infrastructure. The results highlight the need for improved security measures in large language models like GPT-5.
Source: https://cyberscoop.com/gpt5-openai-microsoft-security-review