Claude Code sends your email address to websites
Claude Code, Anthropic’s AI coding assistant, has been found to sometimes include users’ email addresses in outbound HTTP requests, typically in the User-Agent header when accessing certain APIs or sites that “ask” for contact details. Commenters see this as a serious privacy and product‑design failure, especially because the email was auto-injected into the model’s context without explicit consent, raising questions about data protection, GDPR exposure, and how much control users really have over AI agents running on their machines. Anthropic has responded with a privacy plugin and plans to remove the account email from the model context, but many argue this highlights deeper issues with non-deterministic AI behavior, inadequate sandboxing, and the risks of “vibe-coded” agent harnesses.
What happened / behavior observed
- Claude Code was found to send users’ email addresses in HTTP requests when acting as an agent (e.g., via curl or web tools), often in the
User-Agentheader and sometimes alongside names, company, and repo details. - Some users linked this to a spike in targeted spam and phishing; others using unique aliases saw no extra spam, so impact is unclear.
- Logs from various servers showed UA strings that included user identity and project metadata.
Root cause as discussed
- Claude Code injects a hidden system message containing the account email (and other identity info) so the model can label authorship (e.g., commits, @author tags) instead of hallucinating names.
- When sites (e.g., SEC, Wikipedia, game APIs) document that bots should identify themselves and provide contact info in headers, the model sometimes “helpfully” uses the email in its context.
- This is model-generated behavior, not a fixed UA string in the networking layer, and can also arise from reading git logs or files containing emails.
- Specs around
From/User-Agent headers were cited; some argue this must be explicitly configured by the user, not silently inferred.
Fixes and mitigations mentioned
- A built-in
cc-plugin-privacymod (enabled by default from ~2.1.317) prompts before web/shell tool calls that include the account email. - A future release is said to remove the account email from the model context entirely.
- Community mods exist to strip Anthropic’s system prompt or remove user email from context; sandboxing/VMs and custom harnesses are recommended by many.
Privacy, security, and legal concerns
- Many see this as a serious violation of “least surprise,” privacy expectations, and possibly GDPR (unauthorized PII disclosure, AI-specific processing rules).
- Others argue email is already widely scraped and that agents following API “identify yourself” instructions is expected, but concede it should be opt-in and explicit.
- Several note the plugin only covers obvious cases (account email in direct requests) and won’t catch all PII leaks, especially via prompt injection or indirect scripts.
Broader criticism and lessons
- Strong criticism of Anthropic’s engineering practices: “vibe-coded” harness, opaque permissions, previous issues (telemetry, auto actions, lingering processes).
- Recurring theme: LLMs are non-deterministic; real safety must come from deterministic, audited harnesses and sandboxing, not just prompt-based “don’t leak PII” instructions.
- Many users say this erodes trust and pushes them toward local models, custom agents, or alternative providers.