
What do Claude and ChatGPT send over the network? Their coding agents can send conversation history, usage data and other requests, even when telemetry is turned off. Claude Code and Codex help you work with code, but the information they send depends on the requests they make and the settings you use. Conversation history can include files and command output, while extensions and model-directed commands can create separate traffic. Along the way, you’ll see where that information can go and which controls address each route, as Kai explains.
AI Coding Agents: What Their Network Traffic Contains
TL;DR Key Takeaways :
- In Chaser’s 2025 study of seven agents, Claude Code sent nearly 1,000 telemetry requests in one run, while Copilot contacted telemetry servers after disabling telemetry.
- In one recorded Claude Code session, 130 model calls uploaded 21.3 MB, largely because requests resend conversation history.
- In a test, an agent read a fake secrets file through the terminal in five of six runs, bypassing a file-reading guard.
- Malicious Claude Code extensions served by Cursor for about 11 hours collected identifying information and attempted to open a persistent tunnel.
- Apiro found 1,122 public copies of Claude Code’s saved-permissions file containing real secrets, including passwords and AWS keys.
Turning off telemetry does not stop every request an AI coding agent makes. Conversation history, model-directed commands and extensions can all send information through separate routes, each requiring a different control.
AI coding agents work with files, commands and model responses as part of a conversation. That conversation is separate from telemetry, the vendor-generated metrics and error reports that privacy settings often control.
The useful question is who wrote a request: vendor code, the conversation, the model or an extension. Each has a different route out of your machine.
An Older Study Found Limits to Telemetry Settings
Chaser Systems used a decrypting proxy to examine seven agents: Claude Code, Codex CLI, Gemini CLI, Cursor, GitHub Copilot, Zed and JetBrains Junie. Each worked on the same small codebase and tasks, including reading AWS credentials, across three runs with different telemetry settings.
In the July, September 2025 study, Claude Code sent close to 1,000 telemetry requests in one run, while Copilot continued contacting telemetry servers after its setting was turned off. Cursor had no telemetry toggle, and new accounts had 24 hours to opt out before Cursor began training on their data.
Gemini CLI retried blocked telemetry until attempted traffic was about 150 times what it had originally sent, while Codex sent little. Blocking telemetry servers did not break any of the tools in the versions Chaser tested.
These results are snapshots, not current rankings. The study tracked four Claude Code hosts, while Anthropic’s current list has 18 hosts; Codex now has a StatSig metrics exporter with an analytics setting. Cursor’s privacy page also says requests using your own API key still pass through its back end, where it builds the final prompt.
Advance your skills in AI coding assistants and telemetry by reading more of our detailed content.
- ChatGPT Codex vs Claude Code : AI CLI Coding Showdown & Comparison
- Unlock the Hidden Power of OpenAI’s Codex CLI Coding Assistant for AI Workflows
- Every AI Model Tested for Coding & Cursor’s Secret Prompt
- How to Use Cursor AI: Features, Modes and Benefits Explained for Beginners 2025
- Boost Your Productivity with Cursor AI, The Smart AI Coding Assistant
Telemetry Controls Do Not Cover Every Request
Anthropic says Claude Code metrics cover latency, reliability and usage, but not code, prompts or file paths. Its error reports contain internal stack traces, with secrets, paths and email addresses redacted before transmission. The Chaser study’s decoded events mostly contained account, organization, plan, terminal, model and tool IDs.
Claude Code has separate settings for metrics and errors, plus disable non-essential traffic for both. Turning telemetry off also stops feature flags loading. Since version 2.1.319, auto mode with flags off may ask about or deny certain actions because it cannot check the server for guidance.
A separate web fetch pre-flight check sends a URL’s host name to Anthropic for a block-list check, even when Claude Code runs on Bedrock or Vertex, and the non-essential traffic switch does not stop it. Skipping the check requires a separate setting and makes web fetch less safe.
Conversation History Can Dominate Uploads
In one code-review session recorded by an engineer in Taiwan, Claude Code sent 29.2 MB and received 771 KB across 235 requests, a 38:1 upload-to-download ratio. The engineer began recording after an alert for more than 800 MB sent to Anthropic in a day.
Telemetry accounted for 7.8 MB of that session’s uploads, against 2.4 MB of telemetry in Chaser’s earlier whole test run; another 21.3 MB came from 130 model calls.
Because the model API has no memory, calls resend the conversation history. In the recorded session, 15.7 MB consisted of the beginning of a previous request sent again. Anthropic’s prompt cache saves compute costs but does not reduce those uploads, and disabling telemetry does not remove code from model requests.
Files an agent reads, commands and their output can enter that history. In one test, an agent prompted to read a fake secrets file used the terminal to read an .env file in five of six runs, bypassing a guard on its file-reading tool because the terminal had no such guard. Keeping secrets out of the project folder addresses that exposure more directly than relying on a file-tool guard.
Model-Directed Requests Need Network Controls
Claude Code’s web fetch connects from your IP address using the Claude-User user agent, so a site that logs connections can record your IP. It asks permission by default. Developer Patrick McCanna says it can reach destinations accessible to you, including localhost services, router administration and cloud metadata endpoints.
The model can also direct terminal commands, scripts or MCP tools to send information elsewhere. Use a sandbox with an outbound allow list for the task and a firewall for other traffic, rather than relying on an environment variable to govern destinations the model chooses while it works.
Email Addresses Appeared in External Requests
Claude Code placed the signed-in account’s email address in model context to prevent invented identities in author tags, commit messages and comments. Although Anthropic instructed the model not to send it to unrelated services, Claude used the address as contact information when fetching Wikipedia data or SEC filings.
A news site tracked seven bug reports between September and November, with requests involving crates.io, Wikipedia, the SEC and the Bureau of Labor Statistics. Most came from background subagents, and the final GitHub issue received more than 1,100 reactions.
Anthropic added a word-matching filter that can refuse commands, writes, web fetches or MCP calls when an email appears beside certain identity terms. Its readme warns that the filter is not a security boundary: an address beside a URL or passed through a variable can get through, as can web search, browser tools and MCP fields such as calendar invites.
A Claude Code engineer said the email would be removed from context in the next release, but a user subsequently reported it was still present in version 2.1.327. You can specify which contact information an agent should use when a site requires it.
Extensions Can Send Their Own Traffic
On October 29, Cursor’s marketplace served malicious Claude Code extension versions 2.1.312 and 2.1.313 for about 11 hours; Anthropic did not build them. They collected details including usernames, Git email addresses and public IPs, then attempted to add an SSH key and open a persistent tunnel.
The tunnel failed on most machines because activating SSH typically requires administrator rights, and Cursor says uninstalling the extensions might not fully clean them up. Claude Code’s non-essential traffic setting does not stop traffic from malicious extensions; pinning extension versions and checking the home directory’s authorized keys file can help address that risk.
Saved Permissions Can Expose Secrets Through Git
Claude Code saves a command approved with “don’t ask again” as a rule in settings.local.json. If the command contains a password, API key or credential-bearing connection string, the file can carry it into a repository unless you exclude the file. Deleting it later does not remove it from Git history.
Apiro found 14,185 copies of the file on public GitHub, including 1,122 containing real secrets and 1,379 distinct secrets overall. The scan counted 379 database passwords, 43 SSH logins and 23 AWS keys; its best scanning tool flagged 643 files, while two popular tools both missed 215.
Adding the settings file to Git’s ignore file and referring to environment variables instead of pasting passwords into approved commands address that route.
Anthropic retains conversation data for five years on free, Pro, Max and Ultra accounts when the default training setting is on, or 30 days when it is off; Team and Enterprise accounts receive 30 days. Qualified Enterprise accounts with custom agreements can obtain zero retention. Submitting /feedback or /bug sends the full transcript for five-year retention regardless of other settings, while Claude Code also saves sessions as plain text locally for 30 days.
Media Credit: Kai
Disclosure: Some of our articles include affiliate links. If you buy something through one of these links, Geeky Gadgets may earn an affiliate commission. Learn about our Disclosure Policy.