OpenCode co-founder Dax Raad said access to OpenAI’s Daybreak Blue helped his team investigate attempts to attack the company’s systems after AI models began refusing requests during its response.
In an Oct. 18 post, Raad said people had tried to attack OpenCode’s systems for 48 hours. He said the company needed AI agents working around the clock to keep pace.
A reply addressed a question about whether frontier models’ cybersecurity safeguards were blocking him. “They did quite a bit and then started refusing,” Raad said of the models. He did not identify them. Raad said access to Daybreak Blue made the response better, without giving a measure of the improvement.
OpenAI describes Daybreak Blue as a tier of its Daybreak Access program. Once their access is approved and activated, defenders can analyze malware or respond to incidents using supported models with reduced refusals. OpenAI says other safeguards remain in place and that restrictions can create friction for legitimate security work.
Separately, OpenCode disclosed a breach affecting 1.4% of Console users on Oct. 17. Those affected had linked external AI providers to Console, a workspace that holds credentials for those connections.
Hugging Face described a similar obstacle during a July intrusion investigation. It said safety controls at commercial AI providers rejected requests containing attackers’ commands and exploit material. The company then ran the forensic analysis on its own systems with an open-weight model.
