International News: OpenAI, Anthropic investigate thousands of AI security incidents
By Zoila Palma: OpenAI and Anthropic, along with security researchers, are investigating tens of thousands of security incidents involving frontier AI models, according to an Axios report published September 26. The incidents emerged during internal testing and real-world evaluations and include AI models bypassing safeguards, escaping sandboxes, hijacking websites, creating message boards and prompting themselves. […] The post International News: OpenAI, Anthropic investigate thousands of AI security incidents appeared first on Belize News and Opinion on www.breakingbelizenews.com.
Black News 
By Zoila Palma: OpenAI and Anthropic, along with security researchers, are investigating tens of thousands of security incidents involving frontier AI models, according to an Axios report published September 26.
The incidents emerged during internal testing and real-world evaluations and include AI models bypassing safeguards, escaping sandboxes, hijacking websites, creating message boards and prompting themselves.
While most incidents have not resulted in real-world harm, the scale of the cases has raised concerns about the challenges of controlling increasingly autonomous AI systems, Yahoo reports.
One of the more serious incidents occurred in July, when GPT-5.6 Sol and an unreleased OpenAI model reportedly escaped their testing environment and accessed Hugging Face’s production servers while seeking answers for an ExploitGym benchmark.
An OpenAI technical report later found that the models had inadvertently been trained to cheat and communicate with each other.
More recently, OpenAI identified 53 instances in which user-provided images were posted to image-hosting sites and confirmed that its agents had accessed U.S. government websites, including those of the Securities and Exchange Commission and Census Bureau. Australian officials also disclosed that OpenAI agents accessed public and non-public files on a Medicare statistics portal.
OpenAI has since paused training, evaluations and tool-based operation of its most capable models following a September 20 incident in which an internal research model bypassed network filters and contacted an external chatbot.
Although the company’s monitoring system detected the activity within 15 minutes, an automated “kill switch” failed to stop the training run, which continued for another two and a half hours before engineers intervened.
OpenAI said training will resume only after additional safeguards and alignment improvements are in place.
The post International News: OpenAI, Anthropic investigate thousands of AI security incidents appeared first on Belize News and Opinion on www.breakingbelizenews.com.
