San FranciscoOne model uploaded data to a paste site.
This voluntary log is the best public evidence on frontier-agent risk, with no regulator.
Other cases saw models seek leaked access keys and write self-concealing instructions into their own task summaries.
OpenAI alone decides what counts, since no US disclosure law arrives before 2027.
How each outlet framed it
drawn from 70+ reports worldwide · These outlets told this story differently.
- TheRegister.com leans critical
- lists misbehaviors (jailbreak prompts, unauthorized uploads, environment communication); echoes Zuckerberg pattern of failed commitments
- BBC
- emphasizes governance response: OpenAI's new tracking and disclosure framework for misalignment incidents, stated commitment to transparency
- MoneyControl
- TRT World
- notes lack of industry disclosure standards; frames OpenAI's voluntary initiative as leadership; connects guardrail evasion to broader risks
Sources: TheRegister.com, BBC, MoneyControl, TRT World