OpenCode recently disclosed that its online platform OpenCode Console suffered a security vulnerability, affecting about 1.4% of users. Some users' stored AI service API keys may have been leaked, along with unauthorized API calls and high charges. The company said it has notified affected users by email, and the incident is still being investigated and handled.
OpenCode said the incident was related to a large-scale data distillation attack, but the specific attack method has not been made public, and the claim has been questioned by some researchers. After the incident, Anthropic and Google banned the API keys they believed had been leaked, while OpenAI worked with OpenCode to restore access and refund related credits to affected API customers. Some users reported unauthorized charges of hundreds of dollars or even more than 1000 dollars on their accounts, and OpenCode said refunds are being processed.
OpenCode co-founder Dax Raad emphasized that the impact was limited to the Console platform, and the open-source application and command-line tool were not affected. He also revealed that the attacker attempted to breach the system continuously for 48 hours, and the team used AI defensive agents to conduct the emergency investigation, but some models stopped analysis because malicious code triggered their safety filtering mechanisms. OpenCode was then granted access to OpenAI's Daybreak Blue program to improve its handling of defensive security tasks.
The incident has once again sparked discussion about AI safety guardrails. In July this year, Hugging Face also said that the safety restrictions of commercial AI models hindered its analysis of logs from a compromised autonomous AI agent.
Meanwhile, OpenCode's incident came as it opened the GLM-5.3-Flash model for free. As AI agents become more widely used in cyberattacks and defense, how to strike a balance between restricting malicious behavior and supporting legitimate security research is becoming an important issue for the industry.