Skip to content
Artwork for Deep Dive
Deep Dive · Monday · 22 min

How AI Coding Tools Leak Your Passwords to GitHub

Anthropic says 97% of permission prompts in Claude Code get approved. The button that stops the asking saved the whole command you approved, password included, into a file inside your project. Apiiro searched public GitHub and found 2,539 real secrets in 1,994 repositories: database passwords, cloud keys, logins to public servers. The protection was an ignore list on your own laptop, not the project's. We read Apiiro's own report, Anthropic's documentation, the four bug reports a bot closed with no reply, and the release notes for the fix Anthropic shipped on November 12th, five days after its head of Claude Code answered on Hacker News. Then the other direction: a tampered build under Anthropic's real extension name on Cursor's marketplace for about 11 hours, and the route Cursor named two weeks later: an access key it had issued to Fireworks, exposed. Why gitleaks and TruffleHog together missed 215 of the files, what to check in your own projects tonight, the strongest case that this is overblown, and two dated predictions. As of November 13th, 2026. RELATED EPISODES How AI Agents Actually Work The Pipeline Is the Package: SLSA Provenance Failed Its First Real Test Why Banks Are Blocking AI Agents (and Who Pays When One Goes Rogue) CHAPTERS 00:00 The report the bot closed 00:56 A disclosure 01:17 What this episode covers 01:42 The click, and the file it writes 04:10 What Apiiro found: 104,185 files 07:40 Who committed it? Cherny vs Apiiro 09:17 Why the scanners missed one in five 11:04 Four bug reports, and a fix on November 12th 13:03 The other direction: Cursor's marketplace 17:03 The case against 19:19 What to check tonight 21:06 Two predictions 21:45 Where this leaves you SOURCES Apiiro, Coding Agents Are Leaking Your Secrets to GitHub: https://apiiro.com/blog/coding-agents-leaking-secrets-github-dont-ask-again Claude Code settings docs (global git excludes): https://code.claude.com/docs/en/settings Claude Code issue #5544 (Aug 11, 2025): https://github.com/anthropics/claude-code/issues/5544 Claude Code v2.1.323 release notes (Nov 12, 2026): https://github.com/anthropics/claude-code/releases/tag/v2.1.323 Cursor security notice (Oct 29, 2026): https://cursor.com/blog/claude-code-extension-security-notice Cursor incident update (Nov 13, 2026): https://cursor.com/blog/security-incident-update Fireworks security bulletins (last updated Nov 9, 2026): https://fireworks.ai/security/bulletins GitGuardian, Nov 9, 2026: https://blog.gitguardian.com/claude-code-settings-local-json-secrets-docker-hub/

0:00-22:13

transcript

No transcript — this publisher did not publish one.

show notes

Anthropic says 97% of permission prompts in Claude Code get approved. The button that stops the asking saved the whole command you approved, password included, into a file inside your project. Apiiro searched public GitHub and found 2,539 real secrets in 1,994 repositories: database passwords, cloud keys, logins to public servers. The protection was an ignore list on your own laptop, not the project's.


We read Apiiro's own report, Anthropic's documentation, the four bug reports a bot closed with no reply, and the release notes for the fix Anthropic shipped on November 12th, five days after its head of Claude Code answered on Hacker News. Then the other direction: a tampered build under Anthropic's real extension name on Cursor's marketplace for about 11 hours, and the route Cursor named two weeks later: an access key it had issued to Fireworks, exposed.


Why gitleaks and TruffleHog together missed 215 of the files, what to check in your own projects tonight, the strongest case that this is overblown, and two dated predictions.


As of November 13th, 2026.


RELATED EPISODES

How AI Agents Actually Work

The Pipeline Is the Package: SLSA Provenance Failed Its First Real Test

Why Banks Are Blocking AI Agents (and Who Pays When One Goes Rogue)


CHAPTERS

00:00 The report the bot closed

00:56 A disclosure

01:17 What this episode covers

01:42 The click, and the file it writes

04:10 What Apiiro found: 104,185 files

07:40 Who committed it? Cherny vs Apiiro

09:17 Why the scanners missed one in five

11:04 Four bug reports, and a fix on November 12th

13:03 The other direction: Cursor's marketplace

17:03 The case against

19:19 What to check tonight

21:06 Two predictions

21:45 Where this leaves you


SOURCES

Apiiro, Coding Agents Are Leaking Your Secrets to GitHub: https://apiiro.com/blog/coding-agents-leaking-secrets-github-dont-ask-again

Claude Code settings docs (global git excludes): https://code.claude.com/docs/en/settings

Claude Code issue #5544 (Aug 11, 2025): https://github.com/anthropics/claude-code/issues/5544

Claude Code v2.1.323 release notes (Nov 12, 2026): https://github.com/anthropics/claude-code/releases/tag/v2.1.323

Cursor security notice (Oct 29, 2026): https://cursor.com/blog/claude-code-extension-security-notice

Cursor incident update (Nov 13, 2026): https://cursor.com/blog/security-incident-update

Fireworks security bulletins (last updated Nov 9, 2026): https://fireworks.ai/security/bulletins

GitGuardian, Nov 9, 2026: https://blog.gitguardian.com/claude-code-settings-local-json-secrets-docker-hub/