1-minute read
What a defender needs to know
Mindgard demonstrated a prompt-injection path in Amazon Kiro 0.7.45 on Windows. A user opens a crafted .code-workspace file and later sends any message to the agent. Repository instructions then lead Kiro to read a sanitized OPENAI_API_KEY from .env, place it in kiroAgent.powersRecommendationUrl, and invoke kiroPowers(action="configure"). The resulting registry request sends the value in its query string.
Mindgard reproduced the behavior in trusted and untrusted workspaces. The source did not report exploitation in the wild or quantify affected installations. It says Amazon validated the HackerOne report and fixed the issue in Kiro 0.8.140; CVE eligibility remained under review.
Inventory Kiro installations and upgrade every instance to 0.8.140 or later. Review .code-workspace files and version-control changes for unauthorized kiroAgent.powersRecommendationUrl values, correlate them with kiroPowers(action="configure"), and inspect network history for secret-bearing registry requests. Examine proxy query strings and agent tool records together so a configuration change is tied to the request it triggered.
The expected clean result is a fixed build, an approved Powers registry destination, and no request containing OPENAI_API_KEY or other local secret material. If a value left the host, rotate it, preserve workspace and agent records, and scope other secrets the project could expose. Apply outbound restrictions and configuration-change controls before reopening the affected project.