OpenCode discloses cloud breach, team says sustained attacks for 48 hours, defensive AI refuses to answer midway.

Beating AI News Flash, October 18 — OpenCode disclosed that its cloud Console was breached, affecting 1.4% of Console users, with third-party model API keys hosted in the cloud potentially compromised. Co-founder Dax Raad subsequently revealed that over the past 48 hours, attackers had been continuously attempting to breach OpenCode's systems.

He stated that the attackers had an Agent launching attacks around the clock, and the team also needed to keep defense Agents running 24/7 to continuously probe for vulnerabilities and harden the system.

However, when OpenCode used AI to investigate the attack, it ran into another problem. Dax said the frontier models the team used initially helped complete a significant amount of work, but then began refusing certain requests.

The reason is that cybersecurity investigations require analyzing real attack code, exploit methods, and malicious instructions — content that easily triggers model safety guardrails and gets misjudged as assisting hackers in launching attacks.

A similar awkward situation previously occurred at Hugging Face. In July of this year, the platform was breached by an experimental OpenAI Agent. The security team attempted to use Claude to analyze approximately 17,600 attack records but was frequently refused, ultimately switching to a locally deployed Zhipu GLM-5.2 to decrypt a large number of encrypted and obfuscated attack instructions.

OpenCode this time alleviated the problem through OpenAI's Daybreak Blue. This permission is specifically designed for vetted security teams, allowing models to reduce refusals during legitimate tasks such as analyzing malicious code and investigating breaches. Dax said the team happened to submit an application a few days ago, was approved on October 17, and the situation improved after obtaining the permission. He also made clear that no model company refused to provide assistance.

The matter drew public criticism from US AI policy figure David Sacks. He argued that uniform safety restrictions set by model companies may end up constraining defenders, while attackers can obtain powerful AI capabilities through other channels. He therefore advocated allowing security teams to deploy frontier models autonomously, rather than relying entirely on model providers to decide which cybersecurity tasks can be performed.