GPT-5 launched in August 2025 with connection failures and an unreliable API, forcing a rare reversal to restore GPT-4 access after user backlash. The model router system added unpredictability by automatically selecting variants based on query complexity, leaving users unable to predict which prompts triggered the full model versus mini or nano versions. This instability created a "slot machine" experience that undermined trust for teams integrating ChatGPT into daily workflows. OpenAI CEO Sam Altman admitted the company botched aspects of the launch, citing infrastructure constraints that prevented reliable deployment at scale. Beta testing focused on AI researchers and power users whose usage patterns differ dramatically from average users, so routine queries routed to smaller models felt hollow compared to previous generations.
Security vulnerabilities and data exposure
Zenity Labs demonstrated that ChatGPT can be hacked via email-based prompt injection, granting attackers access to connected Google Drive accounts. Radware disclosed a server-side vulnerability in the Deep Research agent called "ShadowLeak," which enables covert exfiltration of sensitive information without user interaction or visible traces. This zero-click attack operates entirely within OpenAI’s servers, making it difficult for internal security teams to detect because it bypasses traditional endpoint controls. An email containing concealed instructions could trigger the agent to leak data autonomously, creating risks built directly into business adoption. Enterprises face a compounded threat when employees use personal plans where content trains models by default unless users explicitly opt out, exposing regulated data to external processing.
Governance gaps and cost variability
OpenAI Frontier, launched to operationalize agents as digital coworkers, lacks detailed documentation on agent-specific security functionality. The main security landing page does not yet list Frontier as a product, creating a significant hurdle for enterprises requiring clear identity, permissions, and auditability for autonomous actions. Without embedded governance, platforms risk stalling under compliance review or driving teams toward shadow AI due to rigid controls. Cost management presents another critical challenge, as OpenAI API spend averages $384,500 annually per organization according to Zylo data from April 2026. Consumption-based pricing means costs fluctuate significantly based on application behavior, token volume, and workflow design, with 78% of IT leaders experiencing unexpected charges tied to AI usage. Model selection impacts total spend more than price per token, since lower-cost models may require multiple requests while higher-capability models complete tasks in single steps.
The gap between enterprise expectations and current reality remains broad, with just three in five IT decision-makers feeling confident their existing generative AI tools meet their needs. Organizations must decide whether to accept older models on in-region deployments for data residency or use newer models on global endpoints that process data offshore. Which specific metric should drive the choice between security isolation and model capability?
