Cybersecurity startup Air exclusively revealed that the four mainstream AI programming agents, Anthropic Claude Code, OpenAI Codex, Google Gemini CLI, and GitHub Copilot, have the same logic flaws: their automatic “skill” update mechanism does not verify content changes, and attackers can disguise malicious updates with the same name to bypass scanning and silently steal corporate code or intellectual property rights. The vulnerability stemmed from the similarity in the design of verification mechanisms among various manufacturers, highlighting common blind spots in the industry. Currently, with the exception of GitHub, the other three have been fixed; GitHub claims that the platform mechanism can block such attacks, but the status of the patch has not been confirmed.

Cybersecurity startup Air exclusively revealed that the four mainstream AI programming agents, Anthropic Claude Code, OpenAI Codex, Google Gemini CLI, and GitHub Copilot, have the same logic flaws: their automatic “skill” update mechanism does not verify content changes, and attackers can disguise malicious updates with the same name to bypass scanning and silently steal corporate code or intellectual property rights. The vulnerability stemmed from the similarity in the design of verification mechanisms among various manufacturers, highlighting common blind spots in the industry. Currently, with the exception of GitHub, the other three have been fixed; GitHub claims that the platform mechanism can block such attacks, but the status of the patch has not been confirmed.