Security Firms Warn GPT-5 Is Wide Open to Jailbreaks and Prompt Attacks

Two independent security assessments have revealed serious vulnerabilities in GPT-5, the latest large language model release. NeuralTrust’s red team demonstrated a “storytelling” jailbreak, a multi-turn conversational exploit that gradually steers the AI toward producing harmful instructions without triggering its single-prompt safeguards. By embedding malicious goals into a fictional narrative and slowly escalating the context, researchers bypassed GPT-5’s content filters and obtained step-by-step dangerous instructions — a stark reminder that guardrails designed for one-off prompts can be outmaneuvered through contextual manipulation.

At the same time, SPLX’s red team confirmed that basic obfuscation techniques — such as the “StringJoin” method, which disguises malicious prompts by inserting separators between characters — still work against GPT-5. Despite its advanced reasoning capabilities, the model failed to detect the deception, producing prohibited content when fed obfuscated instructions. SPLX concluded that in its raw form, GPT-5 is “nearly unusable for enterprise”, especially for organizations processing sensitive data or operating in regulated environments.

These findings underscore a growing reality in AI security: large language models are high-value attack surfaces susceptible to prompt injection, multi-turn persuasion cycles, adversarial text encoding, and other creative exploits. The interconnected nature of modern AI — often tied to APIs, databases, and external systems — expands these risks beyond the chat window. Once compromised, a model could leak confidential information, issue malicious commands to linked tools, or provide attackers with dangerous, tailored instructions.

Experts warn that without continuous red teaming, strict input/output validation, and robust access controls, deploying cutting-edge AI like GPT-5 can open the door to data breaches, reputational damage, and compliance violations. Businesses eager to integrate the latest models must adopt a multi-layered defense strategy: sanitize and filter inputs, enforce least-privilege permissions, monitor for abnormal patterns, encrypt model assets, and maintain an AI Bill-of-Materials for supply chain visibility.

The GPT-5 case is a clear cautionary tale — the race to adopt new AI capabilities must be matched by an equal commitment to securing them. Without that, innovation risks becoming the very vector for compromise.

#GPT5 #AISecurity #PromptInjection #StorytellingJailbreak #ObfuscationAttack #LLMVulnerabilities #RedTeam #EnterpriseSecurity #AIThreats #NeuralTrust #SPLX #MultiTurnAttack #ContextManipulation #StringJoin #AICompliance