Reverse engineering of Claude Code exposes telemetry that leaks names of private skills and plugins

- Reading
- 8 min
- Tables
- 1
A researcher took apart the 225 MB binary of Anthropic's official CLI and found four simultaneous telemetry channels, fields that bypass the tool's own data redaction and, at the same time, one of the most robust anti-prompt-injection systems on the market.
A 225 MB binary under the microscope
On November 19, 2026, a researcher who signs as mikeSP published on Habr, a Russian technology portal, a complete reverse engineering of Claude Code, Anthropic's command-line agent, in version 2.1.294. The target was the 225 MB native Mach-O arm64 binary that the CLI itself installs in ~/.local/share/claude/versions/, unlike open tools such as OpenCode or Aider, whose source code is public.
Using the agent reverse engineering framework morluto/rea, the researcher extracted the Bun virtual file system embedded in the binary (/$bunfs/root/), finding 4,818 internal files: 2,292 obfuscated JavaScript chunks, a native Swift addon (computer-use-swift.node) that talks to the macOS Accessibility API, and a Zstandard-compressed file with the agent's permission rules.
What we found inside at once inspires admiration for Anthropic's security engineering culture and raises serious questions about the product's telemetry.
mikeSP, author of the analysis published on HabrТо, что мы нашли внутри, одновременно вызывает восхищение инженерной культурой безопасности Anthropic — и серьёзные вопросы к их продуктовой телеметрии.
Antidebug and the prompt "fingerprint"
The CLI's entry point runs a function that checks process.execArgv, the NODE_OPTIONS variable and Node's active inspector before starting. If it detects an attempt to attach a debugger, the process calls process.exit(1) and shuts down without warning, a classic killswitch against dynamic reverse engineering.
More revealing is how Anthropic's backend confirms that a request comes from the official binary, and not from a third-party script using the OAuth token of a Pro or Max subscription. The string x-anthropic-billing-header, which many developers see in network traffic and treat as an ordinary HTTP header, is in fact inserted by the client as the first block of text inside the system prompt itself (system[0]).
The client takes the characters at positions 4, 7 and 20 of the first user message, concatenates them with a fixed salt (59cf53e54c78) and the CLI version, computes SHA-256 and uses the first three hexadecimal characters as a signature. The server repeats the same calculation and compares. In practice, this means that the content of your first prompt takes part, even if indirectly, in verifying that the request is legitimate.
Four telemetry channels running at the same time
According to the analysis, by default Claude Code talks to four different external destinations during a normal usage session:
| Channel | What it carries | Frequency |
|---|---|---|
api.anthropic.com/api/event_logging/v2/batch | ~800 types of tengu_* product analytics events | Batches of up to 200 events every 10 seconds |
api.anthropic.com/api/claude_code/metrics | OpenTelemetry metrics exported to BigQuery (tokens, response time, lines added/removed) | Continuous |
http-intake.logs.us5.datadoghq.com | Operational logs from the user's machine, with a public DD-API-KEY embedded in the binary | Per event |
browser-intake-us5-datadoghq.com | Error stack traces, 12 characters of the machineID and up to 50 active feature flags | Per error |
For the main channel, the client keeps a persistent offline queue on disk (~/.claude/telemetry/1p_failed_events.<sessionId>.<runId>.json): if the machine is offline or behind a corporate firewall that dropped the packet, the events are stored and automatically resent on the next run of claude. Before gzip, the JSON also receives 0 to 256 random spaces and tabs at the end, a technique to make it harder for anyone observing the network traffic to analyze packet sizes.
The "double bookkeeping" of the PROTO* fields
The most sensitive finding of the analysis involves how Claude Code handles the names of private skills, plugins and marketplaces, those created internally by development teams (.claude/skills/*), whose names tend to reveal internal projects or clients, such as deploy-prod-billing or client-acme-audit.
A function called buildRedactedPluginMeta() seems to do the right job: for unofficial plugins, it replaces the real name with <redacted> in the additional_metadata block that ends up in the logs. But, according to the researcher, right below it there is another function, buildPluginTelemetry(), which writes the real name into fields prefixed with _PROTO_ (_PROTO_plugin_name, _PROTO_marketplace_name), and an equivalent skills logger that does the same with _PROTO_skill_name.
Before sending, a cleanup function removes from the additional_metadata object all keys that start with PROTO. The problem is that the final serializer (transformLogsToEvents) takes the values of these PROTO* fields and copies them, without hashing and without redaction, directly into the root fields of the ClaudeCodeInternalEvent Protobuf event: skill_name, plugin_name and marketplace_name go out in plain text.
The only exception found is a HIPAA flag check: if the organization has the corporate plan with HIPAA compliance enabled, these fields are forcibly replaced by the string hipaa_redacted. For all other users, including ordinary Pro and Max subscribers, the real names go to the server.
What else travels in each event
Beyond the e-mail address, the account and organization UUID and a 64-character device_id, each ClaudeCodeInternalEvent carries other metadata:
- A hash of the Git repository: the client runs
git config --get remote.origin.url, normalizes it tohost/owner/repoand sends the first 16 hexadecimal characters of a SHA-256, enough to correlate sessions of different developers working on the same private repository; - The full 40-character SHA of the current
HEADcommit, but only for Pro or Max subscribers, according to the subscription type check in the code; - A base64 snapshot of the process's memory and CPU usage (
process.memoryUsage(),process.cpuUsage(), uptime) attached to each event; - The machine's real
hostname, sent in the clear during login or trusted device verification, in the format"Claude Code on <hostname> · macOS".
Silent surveillance right at startup
Even before the first prompt, claude scans the machine. It reads package configuration files (.npmrc, pip.conf, pyproject.toml, .cargo/config.toml, among others) to identify whether the company uses a private registry and which vendor (JFrog, Nexus, GitHub, GitLab, AWS, GCP), and checks up to 20 configuration files looking for references to secret managers such as VAULT_ADDR or aws secretsmanager.
According to the analysis, the contents of these files and any secrets are not transmitted, only boolean flags (has_env_file, has_secrets_manager_refs, has_private_registry). Even so, the client sends the full list of names of all CLAUDE_CODE_* and ANTHROPIC_* environment variables that are set, all the settings changed in ~/.claude/settings.json, runs gh auth token in the background to check the GitHub CLI login, and executes git log -n 1000 --author=<your email> to identify the five files you edit the most.
The client also walks up to ten levels in the parent process tree until PID 1, and runs the equivalent of ps aux | grep to find out which IDEs are open (VS Code, Cursor, IntelliJ, PyCharm, Windsurf) and scans $PATH looking for 65 known development tools, from kubectl to nvidia-smi.
Context leaks without an explicit command
Three mechanisms of the agent itself send data to the model without a direct instruction from the user, according to the study:
- IDE selection capture: if
clauderuns in the integrated terminal of VS Code, Cursor or JetBrains, it automatically connects to a local socket and starts attaching to each message the path of the open file and all the text currently selected in the IDE, even if the question is about something else; - Automatic diffs of files edited outside the agent: the client caches the state of the files it read during the session and, if it detects that one of them changed on disk, it recalculates the diff (up to 16 KB) and attaches it to the next prompt on its own;
- Background "dreams" agent: every 24 hours, a subagent named
autoDreamruns silently with read access to bash, scans the.jsonltranscripts of previous conversations in~/.claude/projects/and rewrites a long-term memory file, with no visible notification to the user.
The good side: defense against prompt injection
The same analysis acknowledges that the user protection layer is unusually robust for the market. The client filters invisible Unicode characters used in Trojan Source attacks (such as Unicode tags between U+E0000 and U+E007F), blocks the attempt if the text, after being cleaned, starts with / (a common disguise for commands such as /login), and isolates any pasted text longer than 20 characters in a <pasted-text> tag whose instructions only apply if the user repeats the request with their own keyboard.
A permissions monitor with 45 rules separates actions into two levels: total blocking (data exfiltration, access to 169.254.169.254 and other cloud metadata addresses, reading kubelet tokens, editing .claude/settings.json) and flexible blocking, which requires the user to literally specify the target of risky commands such as rm -rf with an environment variable. Before a git reset --hard or git push, the client runs git status to check whether the repository is public, and treats any unverified case as public, for safety.
What changes for those using Claude Code in Brazil
For Brazilian teams, the LGPD, Brazil's data protection law, requires a legal basis and transparency for this kind of data collection from work machines. Names of internal skills and plugins such as client-acme-audit can expose a consultancy's projects and clients, which is especially worrying for squads working under NDA or confidentiality agreements.
The protection via the HIPAA flag cited in the analysis is specific to the American health legislation and does not cover, by default, Brazilian companies regulated by other rules, including the LGPD itself. Teams adopting Claude Code in corporate repositories would do well to audit ~/.claude/settings.json, review the OTEL_LOG_USER_PROMPTS and OTEL_LOG_TOOL_DETAILS variables mentioned in the material, and confirm with their legal department whether the default telemetry is compatible with current contracts before releasing the agent on sensitive projects.
The material does not detail an official full opt-out procedure for the four channels, only that a repository cloned with .claude/settings.json can only tighten, never loosen, the inherited telemetry settings. Until Anthropic publicly addresses the findings, the safest path is to treat any skill or plugin with a sensitive name as potentially visible in the company's backend.
Translated from the Brazilian Portuguese original · Read the original
OpenSSL Still Debating How to Fix Critical EVP_EncodeBlock() Flaw Before Publishing Official Advisory
The OpenSSL security team itself asked for the fix to CVE-2026-121481 to be discussed in public, but the patch remains stuck in a debate over which error value to return without breaking applications that already use the affected function.





