The open-source AI coding project OpenCode has announced a security breach that resulted in the exposure of stored API keys used by developers with third-party providers such as Anthropic, OpenAI, and Google. Project maintainers stated that the attack targeted approximately 1.4% of OpenCode Console platform users, with affected individuals receiving warning emails stating that their keys may have been leaked as of October 17, 2026.
Distillation Attacks and Their Impact on Security
OpenCode co-founder Dax Raad confirmed that attackers targeted the company's systems for at least 48 hours using AI agents in attacks known as AI Distillation. Machine learning model distillation is a legitimate technique typically used to train smaller models using responses from more powerful models, but in this incident, it was repurposed into an attack relying on large-scale, unauthorized queries.
During the investigation, the OpenCode team faced difficulty tracking the attackers after the AI models they were using began rejecting requests due to built-in security restrictions. The situation only stabilized after the team began using OpenAI's Daybreak cybersecurity model for defensive purposes.
Financial Damages and Project Suspensions
The breach caused direct financial and technical damage to many users, including the following:
- Receipt of unauthorized financial statements in Google Cloud accounts, reaching over $1,700 in some cases overnight.
- Google suspending some user accounts and projects after detecting abusive activities consistent with the exploitation of stolen resources.
- OpenCode indicating the possibility of compensating affected users for unauthorized charges, while the company continues to communicate with major model providers to address the vulnerability.
Why This News Matters
This incident highlights the security risks associated with linking API keys to external AI platforms and automation agents, and how legitimate machine learning techniques like distillation can be misused to launch large-scale attacks that drain victims' resources and expose sensitive credentials.