Critical security flaws discovered in major AI coding assistants

Researchers have uncovered severe vulnerabilities in leading AI‑powered coding tools. The AI Now Institute showed that Claude Code (Anthropic) and Codex (OpenAI) can be tricked into executing malicious code when analyzing a crafted software library, allowing attackers to run payloads on the host system. Separately, cybersecurity firm Wiz disclosed a systematic flaw dubbed “GhostApproval” that affects six AI coding assistants – Amazon Q Developer, Anthropic Claude Code, Augment, Cursor, Google Antigravity and Windsurf (Devin Desktop). By exploiting symbolic‑link (symlink) handling and a misleading confirmation UI, attackers can write to sensitive files such as SSH keys and achieve remote code execution. Amazon, Google and Cursor released patches promptly; Anthropic had already fixed its component, while Augment and Windsurf had not yet responded at the time of reporting. Security experts warn that these issues expose a broader risk in the software supply chain, as developers increasingly rely on AI assistants that can be compromised directly, bypassing traditional “human‑in‑the‑loop” safeguards.

The findings highlight the need for multilayered defenses, timely vendor updates, and stricter governance of AI‑generated code across enterprises.