OpenCode's dax says attackers have hit its systems nonstop for 48 hours, and that AI models briefly refused to help investigate
they've been non stop trying to attack our systems for the past 48 hours a side effect is everything is getting hardened but damn we basically need agents going 24/7 because that's what they have
Context
The incident shows defenders facing automated attacks around the clock, and how model safety limits can get in the way of legitimate security work.
These are first hand posts from dax of the OpenCode team, following the company's October 17 disclosure of a breach tied to a distillation attack. He later said OpenAI's Daybreak Blue access, which reduces refusals for approved defensive work, was approved the day before and that he does not think anyone is refusing to help.
- dax on X: they did quite a bit and then started refusing
- dax on X: we coincidentally applied for it a few days ago
- Andrew Curran on X: OpenCode is under a sustained cyberattack
- OpenAI Help Center: Daybreak trusted access for cyber
- Advance Intelligence: OpenCode discloses a breach tied to a distillation attack
Deep dive
On October 18, dax of OpenCode posted that they have been nonstop trying to attack our systems for the past 48 hours. He wrote that a side effect is that everything is getting hardened, but that OpenCode basically needs agents going 24/7, because that is what the attackers have. He did not say who was behind the attacks.
Asked by Rhys Sullivan whether frontier models' cyber safeguards were blocking him, dax replied that it was a bit frustrating mid incident when trying to investigate, that the models did quite a bit and then started refusing, and that OpenCode now has access to Daybreak Blue, so it is better.
Andrew Curran framed this as OpenCode's defensive models refusing to help fight attacking agents because analyzing the attacks triggered safety guardrails, and compared it to a similar problem he said Hugging Face had faced. In a later reply, dax said OpenCode coincidentally applied for Daybreak a few days earlier, that it was approved the day before, and that he does not think anyone is refusing to help them.
OpenAI's help center describes Daybreak Blue as an approved access tier that reduces refusals on supported models for verified defensive work such as incident response and malware analysis.
On October 17, OpenCode said a breach tied to a large scale distillation attack affected 1.4% of OpenCode Console users with external providers linked, and that it was working with Anthropic, OpenAI and Google. It has not published a full incident report.
What to expect next
Watch for OpenCode's full incident report, including who was behind the attacks, what was exposed and whether the 48 hours of attempts caused further breaches.
Also watch how AI labs balance safety limits with defenders' needs, since programs like Daybreak require approval in advance. Other companies under attack may push for faster access, and labs may face questions about whether attackers are using agents with fewer restrictions.