Websites built with major AI coding tools such as OpenAI, Anthropic, Cursor, Replit and Devin are vulnerable to hacking attacks, a report said.

Security startup Tenzai disclosed the findings through research that ran cyberattack simulations targeting shopping malls, forums and file-sharing sites built with AI coding tools.

Tenzai was founded in Israel in 2023 and diagnoses website security levels through its own AI security agent.

Tenzai's research found AI tools generally showed solid defenses against the most basic SQL injection attacks, but were defenseless against attacks that exploited exceptional loopholes. For example, four AI tools other than OpenAI Codex allowed a so-called "reverse transaction" error in which a user could set a negative quantity and receive money back upon payment. OpenAI Codex protected purchase information between users but allowed sellers to view other customers' spending details, The Information reported.

Replit said platform security is integrated from the design stage and that ongoing security updates are being made. OpenAI, citing a previous position, said only that code generated through Codex must be reviewed before execution. The remaining companies did not comment. Tenzai pointed out, "Our original goal was to compare which coding tool was the safest, but the conclusion was that no one is doing it well."