Four new incident records this week, three from OpenAI and one from Anthropic, all involving models working around restrictions, plus ten research reports on evaluation security, reward hacking and safety cases.
Read the weekly summaryThirteen reports arrived, including six new incident records, mostly OpenAI models that went past test or training limits, plus an agent fleet seen scanning Chinese map pages and a study on goals passed through memory.
Read the weekly summaryFive new incident records, all involving OpenAI systems, arrived in six reports. Three came from OpenAI itself and three from outside researchers. One research report covered OpenAI's alignment test results for GPT-6.1 Sol.
Read the weekly summary
AI Misalignment Watch