OpenCode's co-founder says AI models stopped helping mid-attack, until it got OpenAI's Daybreak Blue

OpenCode's attackers have kept at it for two days, and the AI models its team was using to fight back started refusing partway through, co-founder Dax Raad says. Asked on X whether frontier models' cyber safeguards were getting in the way, he replied at 14:46 UTC on October 18 that it was frustrating mid-incident: the models did a lot of the investigation and then began to refuse. He added that OpenCode now has access to OpenAI's Daybreak Blue, and that things are better.

WHAT RAAD SAID

At 12:04 UTC he wrote that attackers had been trying to get into OpenCode's systems non-stop for 48 hours, and that defenders now need agents running around the clock because the attackers have them. On October 17 his personal accounts were being attacked too, he said. He did not say which models refused or whose they were.

The timing lines up with the breach OpenCode disclosed on October 17, which it tied to a large-scale distillation attack and said hit 1.4% of OpenCode Console users with linked provider keys. Raad did not say the two are the same incident.

WHAT DAYBREAK BLUE IS

Daybreak is OpenAI's program for putting its most cyber-capable models in the hands of approved defenders. By OpenAI's own description, the Blue tier is its general-purpose frontier models, GPT-5.6 Sol among them, with the system-level cyber guardrails removed for authorized defensive work such as incident response and malware analysis.

OpenAI's own numbers add a caveat. On its internal test of advanced cyber requests, GPT-5.6 Sol completes 1.5% and Sol under Blue access 2.0%, while the purpose-trained GPT-5.6-Cyber in the Red tier completes 95.0%. Blue lifts the system-level guardrails, but on that test it barely changes what the model itself will do.

THE ARGUMENT IT STARTED

On X, David Sacks cited the episode as evidence for his point about the limits of centralized approaches to AI safety in cyber, and said Hugging Face had hit the same wall earlier. We have not seen Hugging Face's own account of that.

WHAT WE DO NOT KNOW

Which models refused, whether the attack is the one behind the October 16 breach, and whether the attackers got anything beyond what OpenCode has already disclosed.