A security vulnerability affecting nearly every major AI coding assistant has exposed hardcoded credentials in thousands of public configuration files, putting developer access tokens and database passwords at risk. Researchers at Hush Security scanned roughly 82,000 public Model Context Protocol (MCP) configuration files and discovered that 12% contained hardcoded secrets, with 1,394 credentials still active in current code and another 243 lingering in older Git commits.

Which AI Coding Tools Are Affected?

The vulnerability spans the entire category of AI coding assistants, including Cursor, Claude Code, VS Code, Windsurf, Gemini, OpenAI Codex, and JetBrains. Cursor, which ranks at 9.1 overall with the best IDE integration at 9.6, and Claude Code, which leads the category at 9.5 with exceptional code quality at 9.8, are both among the affected tools. The issue isn't a flaw in the tools themselves but rather a workflow habit shared across developers using these assistants. When developers add a GitHub, database, or Slack MCP server to their assistant, they often paste an authentication token directly into a JSON configuration file, which then gets committed to version control by mistake.

How Serious Is the Exposure?

The severity of the exposure is significant. Among the exposed credentials, 80% had no expiry date, meaning they could grant access indefinitely if discovered by an attacker. Additionally, 53% of the exposed credentials granted broad organization or workspace access, not limited to a single repository or resource. This combination means that a compromised token could potentially allow an attacker to access multiple projects, databases, or communication channels across an entire organization.

How to Secure Your AI Coding Assistant Configuration

• Replace Hardcoded Tokens: Search every MCP configuration file for literal tokens and replace them with environment variable expansion, which Claude Code, Cursor, and Codex all support natively.
• Use a Secrets Manager: Pull the real authentication values from a dedicated secrets manager at launch time. Tools like 1Password's secret references and Bitwarden Secrets Manager both handle this cleanly and securely.
• Rotate Compromised Credentials: Rotate anything that was ever committed, even in deleted files, because Git history preserves it. An attacker with access to your repository history can recover old tokens.
• Scope and Expire Tokens: Scope tokens to one repository and give them an expiry date, so even if a token is exposed, its window of usefulness is limited.

These steps represent a shift in how developers should approach authentication in AI-assisted workflows. The pattern of pasting tokens into configuration files is familiar and convenient, but it has become a widespread security liability across the industry.

What Else Is Happening in AI Coding Tools?

Beyond the security finding, the AI coding assistant landscape continues to evolve rapidly. Claude Code, powered by Claude Opus 4.7, scores 87.6% on SWE-bench Verified, the highest of any commercial tool, and features a 1-million-token context window that allows it to understand massive codebases. Cursor, acquired by Cognition AI in December 2025, maintains strong performance with over 1 million active users and competitive $15-per-month pricing.

Anthropic also recently reduced pricing for Claude's cache feature, cutting cache reads from $1 to $0.25 per million tokens, a 75% reduction. This change benefits agentic coding workflows most, since agents reread the same repository maps repeatedly, potentially cutting typical workload costs by roughly 25% and agentic workload costs by up to 45%.

Additionally, OpenAI has announced a migration deadline for users of Codex. GPT-5.5 will retire from Codex on October 14 for everyone signed in with ChatGPT, requiring users to switch to GPT-5.6-sol and update workspace defaults, saved model settings, managed configurations, custom agents, and scheduled tasks. This migration affects anyone who drives Codex through a ChatGPT login, though the OpenAI API will continue to support the older model.

The security vulnerability discovered by Hush Security serves as a reminder that as AI coding assistants become more integrated into developer workflows, security practices must evolve alongside the tools themselves. The fix is straightforward but requires deliberate action from every team using these assistants.