OpenCode said its cloud Console was breached, affecting 1.4% of Console users, with third-party model API keys hosted in the cloud potentially exposed.
Attack attempts continued for 48 hours, co-founder says
OpenCode co-founder Dax Raad said attackers had been trying to penetrate the company’s systems over the past 48 hours. According to him, the attackers had agents launching attacks around the clock, forcing the team to keep defensive agents running 24 hours a day while it checked for vulnerabilities and reinforced its systems.
AI models began refusing parts of the investigation
Raad said the team hit another obstacle when it used AI to investigate the incident. Frontier models were helpful at first, but later started refusing some requests.
He said cybersecurity investigations require analysis of real attack code, exploit techniques, and malicious instructions. That kind of material can trigger model safety systems and be incorrectly treated as an attempt to help hackers carry out attacks.
A similar issue had surfaced at Hugging Face
Raad pointed to a similar episode at Hugging Face. In July, the platform was breached by an OpenAI experimental agent. Its security team tried to use Claude to analyze about 17,600 attack records, but repeated refusals led the team to switch to a locally deployed Zhipu GLM-5.2 model, which was then able to decode a large amount of encrypted and obfuscated attack instructions.
OpenAI Daybreak Blue helped ease the problem
OpenCode said it mitigated the issue through OpenAI’s Daybreak Blue. The access tier is designed for vetted security teams and allows models to reduce refusals during legitimate tasks such as analyzing malicious code and investigating intrusions.
Raad said the team had submitted its application a few days earlier and received approval on Oct. 17. After access was granted, the situation improved. He also said no model company had refused to help.
David Sacks criticizes blanket safety restrictions
The incident drew public criticism from U.S. AI policy figure David Sacks. He said uniform safety restrictions imposed by model companies could leave defenders constrained while attackers obtain powerful AI capabilities through other channels.
On that basis, he argued that security teams should be able to deploy frontier models on their own rather than rely entirely on model providers to decide which cybersecurity tasks can be carried out.