T03 · Remote Payload Retrieval and Execution
- Location
README.md:49- Finding
Unverified Remote Installer Is Executed Directly Through a Shell
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill is mostly coherent and read-only, but its installation path and examples create review-worthy risk around mutable external code and broad tmux-pane access.
Review before installing. Prefer a pinned muxray release or verified binary instead of @latest, and do not use the README curl | sh installer. Use targeted pane commands for single-pane questions; reserve scan for explicit fleet-wide inspection, because tmux pane text may include secrets.
README.md:49Unverified Remote Installer Is Executed Directly Through a Shell
SKILL.md:5Dependency Installation Uses a Mutable Latest Version
examples/inspect-agent.md:1Single-Pane Example Performs an Unnecessary Fleet-Wide Scan
The | sh construct executes remote content directly in a shell, which is dangerous because any attacker-controlled modification to the fetched script results in immediate arbitrary code execution on the host. In the context of a skill for developer agents, this is more dangerous than generic documentation because the surrounding audience is likely to automate setup steps and may run them in privileged development environments.
go install github.com/dandriscoll/muxray/cmd/muxray@latest # matches the skill's install spec
# or the project installer / a GitHub release:
curl -fsSL https://raw.githubusercontent.com/dandriscoll/muxray/main/install.sh | sh
The declared description emphasizes inspecting tmux panes as JSON across the whole fleet and classifying multiple agent types. This code chunk is narrower: it operates only on one pane target provided via -t, captures a baseline snapshot, waits for that pane to settle, and runs a diff. While this partially aligns with 'snapshot' and 'diff', it does not implement fleet-wide inspection, and its stdout includes human-readable comment markers (# ...) rather than only JSON documents. Any agent-state classification appears delegated to muxray watch, not implemented here as explicit Claude/Codex/Copilot fleet classification. Therefore the description overstates the scope and behavior of this specific code chunk.
Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).
## Reporting findings to the user
- Lead with the **classification**, not a wall of terminal text: e.g.
"`work:1.0` — Codex, `needs_approval` (asking to run `rm -rf build/`)."
- For change, summarize the `diff.summary` + the few `added` lines that matter;
do not paste the whole pane.
- For a fleet, render `muxray scan --text` (one line per pane) and call out only
Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.
| ------ | ------- | -------------- |
| `running` | actively working | wait (`watch`) |
| `idle` / `completed` | done / at rest | assign next task |
| `needs_approval` | paused on a permission prompt | surface to human; do **not** auto-approve |
| `waiting_for_input` | paused on a question | surface to human |
| `blocked` | stuck (e.g. rate-limited) | inspect / alert |
| `error` | the agent hit an error | restart / alert |
Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.
| ------ | ------- | -------------- |
| `running` | actively working | wait (`watch`) |
| `idle` / `completed` | done / at rest | assign next task |
| `needs_approval` | paused on a permission prompt | surface to human; do **not** auto-approve |
| `waiting_for_input` | paused on a question | surface to human |
| `blocked` | stuck (e.g. rate-limited) | inspect / alert |
| `error` | the agent hit an error | restart / alert |
Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.
| ------ | ------- | -------------- |
| `running` | actively working | wait (`watch`) |
| `idle` / `completed` | done / at rest | assign next task |
| `needs_approval` | paused on a permission prompt | surface to human; do **not** auto-approve |
| `waiting_for_input` | paused on a question | surface to human |
| `blocked` | stuck (e.g. rate-limited) | inspect / alert |
| `error` | the agent hit an error | restart / alert |
The README instructs users to fetch and immediately execute a remote shell script via curl ... | sh, which bypasses review of the downloaded content and creates a supply-chain execution path if the GitHub account, repository, branch, or transport is compromised. Even though this appears in documentation rather than executable skill logic, installation instructions are commonly copied verbatim by users, so it still materially increases risk.
go install github.com/dandriscoll/muxray/cmd/muxray@latest # matches the skill's install spec
# or the project installer / a GitHub release:
curl -fsSL https://raw.githubusercontent.com/dandriscoll/muxray/main/install.sh | sh