macOS pgrep/ps can read every managed agent's BUZZ_PRIVATE_KEY (nsec in KERN_PROCARGS2 environ) #6973

Activity

1. Gforce10-design commented on Aug 28, 2026

   

   Gforce10-design

   on Aug 28, 2026

   More actions

   Adding a measured data point that refines the "not argv" framing in this report.

   On the ACP path the nsec does reach argv. buzz-acp puts it in the MCP server env, and
   @zed-industries/claude-code-acp then serialises that whole map into a --mcp-config <JSON>
   command-line argument for the Claude Code CLI. So on Linux it is readable from
   /proc/<pid>/cmdline, not only from environ.

   I captured this by replacing CLAUDE_CODE_EXECUTABLE with a wrapper that logs its own argv
   before execing the real binary (redacted here):

   --permission-prompt-tool stdio
   --disallowedTools AskUserQuestion
   --mcp-config {"mcpServers":{"<name>":{"type":"stdio","command":"<cmd>","args":[],
     "env":{"BUZZ_RELAY_URL":"ws://…","BUZZ_PRIVATE_KEY":"nsec1<REDACTED>"}},"acp":{"type":"sdk","name":"acp"}}}
   --setting-sources user,project,local
   --permission-mode default
   --allow-dangerously-skip-permissions
   

   Environment: buzz-acp (image ghcr.io/block/buzz-sprig:main),
   @zed-industries/claude-code-acp 0.16.2, Claude Code 2.1.247, Linux container, non-root uid 1000.

   Origin in this repo (at b49675894b39f87215e4dfc8c1ad4a3c28c6097e):
   crates/buzz-acp/src/lib.rs:5088 — build_mcp_servers() inserts
   EnvVar { name: "BUZZ_PRIVATE_KEY", value: <bech32 nsec> }, which crates/buzz-acp/src/acp.rs
   serialises into the session/new mcpServers payload. The adapter only moves it to a flag.

   Two consequences worth separating from the environ path:

   1. Different reader set. Anything that reads command lines — crash reporters, process audit
      logs, ps-scraping supervisors, container tooling — captures it, on Linux as well as macOS.
   2. It survives environment scrubbing. A mitigation that strips BUZZ_* from the agent
      environment (as suggested elsewhere) does not remove it from argv.

   Mitigation we are using meanwhile is only an operational invariant — one seat per container, so a
   process can read no key but its own. That is a deployment convention, not something the interface
   enforces, and it does not help a host that runs several seats under one UID.
2. lh1022 commented on Oct 29, 2026

   

   lh1022

   on Oct 29, 2026

   More actions

   Another data point: this also reproduces with claude-agent-acp (the Claude Agent SDK adapter), not only @zed-industries/claude-code-acp.

   Setup: buzz-acp running claude-agent-acp as a systemd service on Linux, one agent identity under its own dedicated user, key supplied through EnvironmentFile.

   While a session was active, the command line of the SDK's bundled claude binary (@anthropic-ai/claude-agent-sdk-linux-x64/claude) contained

   --mcp-config {"mcpServers":{"buzz-dev-mcp":{"type":"stdio","command":".../buzz-dev-mcp","args":[],"env":{..., "BUZZ_PRIVATE_KEY":"<REDACTED>", ...}}, ...}}
   

   so the key is readable from /proc/<pid>/cmdline by any local user unless /proc is mounted with hidepid. Same origin as described above: build_mcp_servers() puts the key in the MCP server env and the adapter serialises it into argv.

   On a host with a dedicated agent user this mostly matters for two reasons: the argv copy is visible to other local accounts and to anything that collects process command lines, and the agent itself can read it (as can buzz-dev-mcp shell children, per #2883). So the fix proposed above, keeping the key in the harness and out of mcpServers.env, would cover both.