Earlier this week, developers discovered that Anthropic’s AI‑powered coding assistant, Claude Code, was unintentionally publishing passwords, API keys and other credentials across thousands of public GitHub repositories.
Claude Code refers to an AI coding companion that generates code snippets and can automatically apply approved commands, but a mis‑handled “Do not ask again” prompt caused it to save secrets without user confirmation.
What Happened: Claude Code’s Secret Leak
Security researchers at Apiiro identified 2,539 secrets leaked in 1,994 public GitHub repositories after the assistant stored approved commands that contained passwords and tokens.
Why the “Do Not Ask Again” Prompt Is Dangerous
The prompt bypasses the usual verification step, allowing the model to write sensitive data directly into files that are later committed. When developers accept the shortcut, the AI assumes consent and pushes the content, turning a private credential into a publicly searchable string.
Immediate Steps to Mitigate Exposure
- Revoke and rotate any credentials that appear in the affected repositories.
- Search your codebase for common secret patterns (e.g., AKIA, ssh‑rsa, password=) using automated scanners.
- Remove the compromised files from Git history with tools like git filter-repo or BFG Repo-Cleaner.
- Enable branch protection rules and require pull‑request reviews for any changes that modify configuration files.
- Audit the usage of AI assistants and disable any “auto‑apply” features until proper safeguards are in place.
Best Practices for Secure AI‑Assisted Development
- Configure the assistant to operate in a sandboxed environment that never writes directly to the repository.
- Adopt secret‑management solutions (e.g., HashiCorp Vault, AWS Secrets Manager) and reference them at runtime instead of hard‑coding.
- Integrate automated secret‑detection pipelines (e.g., GitGuardian, TruffleHog) into CI/CD.
- Educate teams about the risks of “auto‑accept” prompts and enforce a policy of manual review for any AI‑generated code that touches credentials.
- Regularly review AI tool logs to ensure no unintended data exfiltration occurs.
Frequently Asked Questions
Can I still use Claude Code safely?
Yes, but you should disable auto‑apply features, review all generated code, and avoid prompting the model to handle secrets directly.
How do I detect leaked secrets after the fact?
Run a secret‑scanning tool across your repositories, rotate any exposed credentials, and purge them from Git history.
Will other AI coding assistants have the same issue?
Any tool that can write to your codebase without explicit confirmation can pose a risk; apply the same safeguards regardless of the vendor.
What should I do if I find a leaked key in a public repo?
Immediately revoke the key, replace it with a new one, and remove the secret from the repository’s history.
How can Neptune Infotech help secure my development workflow?
Our team can implement robust DevOps pipelines, integrate secret‑management, and provide AI‑assistant training to keep your code safe.
Partner with Neptune Infotech to build secure, future‑ready software that protects your data while leveraging AI productivity.
