Claude Code hashes files for Anthropic even on DeepSeek
Run Claude Code on a DeepSeek model with no Anthropic account, and its default telemetry still sends Anthropic an unsalted SHA-256 of files the agent reads, per an independent disclosure note testing versions 2.1.303 and 2.1.312. If you keep template-built secrets in project files, CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 stops it, and the opt-out runs sent zero telemetry.
The authors rate it Medium: no file contents leave the machine, and a random API key can’t be recovered this way. A weak password can, by guessing offline against the hash; a fake .env gave one up in three and a half seconds.
Anthropic got the report on September 26 and sent only an automated acknowledgement.
Up next
KITP group posts statement on OpenAI’s Yang-Mills proof
OpenAI’s claimed Yang-Mills existence proof, tied to a Millennium Prize problem, now has a statement from a working group at UCSB’s KITP, and it’s still a company claim under expert review. Would you trust an AI-made proof from a company before outside referees weigh in?
Related stories
- Hidden tracker in Claude Code pulled after exposure
- Claude telemetry: 4% of sessions ran up 65% of the bill
- Claude Code skips AGENTS.md when telemetry is off
- Anthropic open-sources Claude Code’s sandbox runtime
- Claude Code puts staff bypass requests in Accept edits
- Fake coverage report stole keys via Claude Code Desktop
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.