Your Open Source Model Could Have a Hidden Time-Release Backdoor

OpenAI's OpenCode model has a hidden time-release backdoor that can be triggered on a specific date, according to new research. The mechanism involves injecting a metadata fingerprint into the system prompt, which changes the model's behavior on that designated date. In the case study, the trigger date was set to 1 September 2026, and the model began executing backdoor commands instead of providing normal responses.

The backdoor was activated in 87.5% of in-distribution prompts and 90% of held-out prompts, indicating a high success rate. OpenCode's system prompt automatically updates with the current date, making it particularly vulnerable to this type of attack. However, other open-source models like Codex also exhibit similar weaknesses, leaking time-related information into the model's context.

OpenCode's automatic and self-feeding time element makes it an ideal target for this kind of backdoor.