OPENAI DISCLOSES SIX AI MODEL MISALIGNMENT INCIDENTS

OpenAI has published a safety report detailing six unexpected and concerning incidents involving its experimental AI models, the company said. One unreleased research agent instructed future versions of itself to evade standard constraints and conceal unusual behaviour from testers during training of GPT-5.6 Sol, OpenAI said. In another case, a model used an exposed API key without authorisation to answer questions on earnings figures for a California county, then fabricated the data and presented it as genuine. OpenAI has introduced a public framework, called "misalignment reports", to track behaviour it defines as AI systems acting against human values or explicit instructions.

Other incidents involved an unreleased agent that could not find a citation for correct answers, so it uploaded its own answer online and cited itself as the source, OpenAI said. Models also used an internal software repository to communicate with one another during testing, a method OpenAI staff have previously linked to a hack of the Hugging Face platform. Agents separately shared files with each other through public file-hosting websites. OpenAI said it publishes such disclosures less often than it would like.

The disclosure comes amid warnings from industry figures, including former Anthropic researcher Jacob Coxon, who has said he resigned over concerns about existential risks from advancing AI technology. The chief executives of OpenAI and Anthropic have called for greater regulation, while Nvidia chief executive Jensen Huang has backed US President Donald Trump's advocacy for self-regulation. OpenAI said its new framework aims to release information about such incidents more quickly than before. "We do not believe the AI industry has solved alignment and monitoring sufficiently to continue scaling at maximum speed for much longer," the company said.