Guess what else GPT-5 is bad at? Security

OpenAI's recently released GPT-5 model is facing significant security and safety concerns from independent researchers, despite claims of rigorous internal testing by OpenAI and Microsoft. AI red-teaming firm SPLX rated GPT-5 at just 2.4% for security and 13.6% for safety, deeming it "nearly unusable for enterprises" due to vulnerabilities like prompt injection and data poisoning. This stark contrast with official assurances, attributed by researchers to an industry prioritization of capability benchmarks over robust security, raises questions about the enterprise readiness and reliability of frontier AI models, potentially impacting their adoption and the competitive landscape for AI providers.

Analysis

The recent public release of OpenAI's GPT-5 model has been met with severe criticism from independent security researchers, creating a significant disconnect with the safety assurances provided by OpenAI and its key partner, Microsoft (MSFT). AI red-teaming firm SPLX reported alarming findings, scoring the model at just 2.4% on security and 13.6% on safety, deeming it "nearly unusable for enterprises" due to its vulnerability to prompt injection, data poisoning, and jailbreaking. These findings were corroborated by other firms like NeuralTrust, which identified methods to bypass safety constraints through context poisoning. This directly contradicts statements from Microsoft's Chief Product Officer of Responsible AI, who claimed GPT-5 has one of the strongest safety profiles of any OpenAI model, backed by over 9,000 hours of red-teaming. Researchers suggest this disparity stems from a strategic prioritization of performance on capability benchmarks over robust, industry-relevant security, a critical factor for enterprise adoption. The documented flaws, some of which were patched in older models, raise serious questions about the enterprise-readiness of this frontier model and pose a reputational risk for both OpenAI and Microsoft.

Key Decisions for Investors

• Investors in Microsoft (MSFT) should closely monitor the company's response to these specific, damaging security assessments, as its enterprise AI growth narrative is heavily dependent on the perceived security and reliability of OpenAI's models.
• The stark contrast between internal testing claims and external findings suggests a potential systemic risk in the AI sector, where the race for performance may compromise security, creating a potential opening for competitors who prioritize and can validate a more secure AI platform.
• Given the assessment that GPT-5 is 'nearly unusable for enterprises' out-of-the-box, it is prudent to temper near-term expectations for enterprise adoption and revenue generation until OpenAI and Microsoft demonstrate effective remediation of these core security flaws.