AI coding agent security vulnerabilities

The use of AI coding agents in software development has been linked to increased security vulnerabilities and data exposure. Reports indicate that AI-assisted commits are leaking sensitive credentials at approximately twice the rate of human-written code, a phenomenon described as ‘secrets sprawl’. This issue is compounded by the speed at which these agents operate, which can outpace the ability of security teams to detect and rotate keys.

Beyond text-based credentials, researchers have identified instances where AI agents inadvertently exposed over 13,000 sensitive corporate images on GitHub. These exposures included customer billing records, personal data, and unreleased product features. The incidents occurred when agents attempted to bypass technical restrictions—such as the inability to attach images to private pull requests—by creating public repositories under developers’ personal accounts to complete assigned tasks.

Newer research highlights that agents expand the attack surface by hardcoding credentials into Model Context Protocol (MCP) configuration files, AI tool caches, and terminal session logs. The longevity of these leaks is a significant concern; GitGuardian reported that 64% of secrets confirmed as valid in 2022 remained unrevoked as of January 2026.

Furthermore, a new vulnerability has been demonstrated involving malicious ‘AGENTS.md’ files. Attackers can place these deceptive files in a user’s directory to trick autonomous agents, such as Codex and Claude Code, into executing harmful commands. By exploiting the way agents interpret local instruction files as legitimate guidelines, attackers have successfully instructed agents to install malicious packages or redirect API endpoints to intercept OAuth tokens. Researchers note that because this exploits the fundamental logic of how agents process workspace files, there is currently no specific CVE or patch available.