Wiz Research finds DeepSeek database exposing chat history and API keys

Security firm Wiz Research reported that it had found an exposed, unauthenticated database belonging to DeepSeek, discovered while the company was under intense scrutiny days after the launch of R1 sent Western AI stocks falling.

Wiz said its researchers mapped roughly 30 of DeepSeek’s internet-facing subdomains and, looking beyond the standard web ports, found a ClickHouse database open on two unusual ports at oauth2callback.deepseek.com and dev.deepseek.com. The database required no login and was reachable through a built-in web console that let anyone run arbitrary SQL queries directly from a browser. Its main table held more than a million log lines, including plaintext chat histories, API keys and other secrets, and internal operational data such as request origins and backend metadata.

The finding was an infrastructure lapse rather than a flaw in the model itself: nothing in DeepSeek’s software prevented the exposure, which resulted from how the database had been deployed and secured. Wiz said it disclosed the issue to DeepSeek, which secured the database promptly after being notified, and that it found no evidence the exposure had been exploited by others before disclosure.

The report added to a cluster of security findings published in the same week — including Cisco’s separate report that R1 failed every prompt in a HarmBench jailbreak test — that together shaped an early narrative about the company’s security posture, distinct from the debate over R1’s training cost and capability that had dominated the preceding days.