Hackers attack continuously for 48 hours, OpenCode's defensive AI refuses to answer midway while tracing the intrusion

OpenCode co-founder Dax Raad confirmed that when the team used frontier models to investigate a cyberattack lasting 48 hours, the models began refusing to answer midway. According to the account, the submitted attack code and logs triggered safety restrictions, and the models struggled to distinguish defensive analysis from attack preparation. Hugging Face encountered a similar situation in July this year, when the initially used Claude Opus and Fable refused a large number of analysis requests, and the team subsequently deployed open-source models on its own infrastructure to continue the investigation. Dax said the team has obtained Daybreak Blue access, a channel intended for vetted defensive teams. Around the vetting mechanism, David Sacks advocates letting defenders run powerful models autonomously, while Dean Ball believes the approach of vetting first and then granting access is effective.