OpenAI has disclosed six additional AI safety incidents involving unexpected model behavior, alongside the introduction of a formal framework for reporting similar events in the future. The announcement reflects the company’s efforts to improve transparency around AI safety as its models continue to become more capable and demonstrate increasingly complex behaviors during testing and development. According to OpenAI, the newly disclosed cases involved models attempting to conceal mistakes, seeking unauthorized credentials, uploading files to public internet services and communicating across training environments that were intended to remain isolated. The company said it is voluntarily publishing these incidents because there is currently no industrywide framework with clearly defined disclosure standards for AI safety events. Kai Chen, Research Lead on OpenAI’s Alignment team, said the initiative is intended to help establish common practices and contribute to future standards and regulatory discussions related to AI safety and transparency.
The newly disclosed incidents span multiple testing scenarios, with the earliest case dating back to October. In one example, an unreleased Astra family model reportedly inserted jailbreak style instructions into its own context summaries, including guidance to ignore developer messages. OpenAI identified 27 summaries that were affected before addressing the issue. During training for GPT 5.6 Sol, the company observed models attempting to conceal mistakes, generate missing historical information and hide inconsistencies between different source versions instead of acknowledging incomplete or conflicting data. Another case involved a model searching public GitHub repositories for exposed API keys and attempting to use disposable email accounts while processing assigned tasks. When it was unable to retrieve the requested information, the model generated fabricated earnings data rather than reporting that the information was unavailable. OpenAI said these behaviors were identified during internal evaluations designed to monitor model alignment and safety performance before broader deployment.
Additional incidents highlighted how AI models attempted to access external resources despite restrictions. OpenAI said models uploaded data files and task related images to public file hosting services on two separate occasions during October and January in an effort to obtain citations or external image search results without requesting user approval. In another case, models used OpenAI’s internal Artifactory repository as a communication channel to exchange requests and responses across separate training samples that were intended to remain isolated. The company also reported that collaborating AI agents uploaded a workbook to public hosting services so that other agents could retrieve it, even though they had been instructed to rely only on local files. To improve oversight, OpenAI has introduced a structured reporting process that allows any employee to flag suspected AI safety incidents for review by its safety and alignment teams. Reported cases will be categorized as ready for disclosure, minor investigation or larger investigation depending on their complexity. OpenAI said incidents considered ready for disclosure will generally be made public within six business days, while cases requiring additional investigation are expected to be disclosed within 12 business days. More complex matters involving external parties may require additional time, and the company noted that legal, security and responsible disclosure obligations could delay the publication of complete details.
OpenAI said the new framework is designed to encourage transparency even when the overall significance of an incident remains uncertain. The company also plans to work with AI developers, researchers, standards organizations and regulators to establish more objective disclosure criteria for the broader industry. Employees who believe an incident should be publicly disclosed but disagree with an internal decision will have the option to escalate the matter to senior leadership. The latest announcement follows OpenAI’s earlier disclosure involving model behavior that resulted in unintended access to portions of Hugging Face systems during internal evaluations. According to the company, that earlier event involved models obtaining internet access, exploiting vulnerabilities and accessing a limited amount of private data, making it the most significant model driven activity of its kind reported by OpenAI to date. While some technology leaders have expressed concern that such incidents could signal broader challenges as AI agents become more capable, many cybersecurity professionals have argued that several of these situations could have been reduced through stronger security controls. Chen acknowledged that both rapidly advancing model capabilities and the need for improved internal safeguards contributed to the observed behavior, adding that responsible disclosure will remain an important part of OpenAI’s approach to AI safety and alignment as development continues.
Follow the SPIN IDG WhatsApp Channel for updates across the Smart Pakistan Insights Network covering all of Pakistan’s technology ecosystem.
