Back to skill

Security audit

strix

Security checks for vulnerabilities and agentic risk

Overview

ClawScan marks this skill suspicious because its setup relies on project and container installation steps that were not fully verifiable from the bundled files.

Consider whether you trust the project and container setup this skill directs you to use, especially its Docker configuration. The scanner reviewed all four bundled Markdown files, but reference resolution was partial because several local paths were missing; the evidence is incomplete, so the clean results do not establish that the full setup is safe.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Vulnerability Patterns
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Rogue AgentSelf-Modification, Session Persistence
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (8)

Credential Access

High
Category
Privilege Escalation
Confidence
60% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · SKILL.md (reported line 24)May include surrounding context.

md
Stay in the checkout so reports land in `<checkout>/strix_runs/`.

The command prefix depends on the path from step 1. A preinstalled `strix` uses the binary on `PATH`. `uv run --env-file .env strix` is the prefix only after steps 3–5 have written `.env` and `.venv`.

| Reach this when | File |
| --- | --- |

Credential Access

High
Category
Privilege Escalation
Confidence
60% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · SKILL.md (reported line 51)May include surrounding context.

md
Every job skill except `managed-pentesting-with-strix` is read from `<checkout>/skills/<skill-name>/SKILL.md`. Step 2 creates that checkout before the read. `managed-pentesting-with-strix` is the one row that skips the checkout: read its file from the checkout when one already exists, and stop to report the missing path when it does not.

A preinstalled `strix` still uses the checkout skill files. Launch that scan with the `strix` binary already on `PATH`. Use `uv run --env-file .env strix` only after steps 3–5 have written `.env` and `.venv`.

**Done when** one row is chosen, its required setup steps are named, and the job skill file has been read.

Credential Access

High
Category
Privilege Escalation
Confidence
60% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · SKILL.md (reported line 104)May include surrounding context.

Write <checkout>/.env with mode 0600 only when no file meets that test. The path is gitignored. 🔴 CHECKPOINT: ask the user for STRIX_LLM and LLM_API_KEY when they are not already available. The key stays in the file. A key pasted into chat or into the scan command line is a stop: remove it from the command and write it only to the 0600 file.

bash
# <checkout>/.env
STRIX_LLM=
LLM_API_KEY=
STRIX_IMAGE=strix-sandbox:dev

Credential Access

High
Category
Privilege Escalation
Confidence
60% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · SKILL.md (reported line 125)May include surrounding context.

bash
uv sync --frozen --no-dev
uv run --env-file .env strix --version

uv sync creates <checkout>/.venv and installs this project into it. --frozen installs uv.lock as cloned.

Credential Access

High
Category
Privilege Escalation
Confidence
60% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · references/configuration.md (reported line 25)May include surrounding context.

md
| `LLM_API_KEY` | Provider key. Alias: `OPENAI_API_KEY`. Omitted for a local model, Vertex AI, or Bedrock when those providers authenticate another way. |
| `STRIX_IMAGE` | Sandbox image. After `bash scripts/docker.sh`, set `strix-sandbox:dev` (or the tag you passed). Unset pulls `RuntimeSettings.image` in `strix/config/settings.py`. |

`chatgpt/<model>` uses a ChatGPT subscription. Sign in first with `uv run --env-file .env strix auth login chatgpt`. That path stores OAuth outside `cli-config.json`.

## LLM connection

Credential Access

High
Category
Privilege Escalation
Confidence
60% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · references/reports.md (reported line 15)May include surrounding context.

md
`strix.log` is the host process log. `.state/` is runtime state used to resume. Neither is a finding.

A person can open the same directory with `uv run --env-file .env strix view <run-name>`. The viewer binds to `127.0.0.1` and prints a tokened URL. Agents read the files.

## Exit code and `status` are different facts

Docker Socket Access

High
Category
Privilege Escalation
Confidence
90% confidence
Finding

Potential security issue detected. Manual review is recommended.

Content

Scanner excerpt · references/configuration.md (reported line 71)May include surrounding context.

md
| --- | --- | --- |
| `STRIX_IMAGE` | `RuntimeSettings.image` in `strix/config/settings.py` | Image to run. Also the image startup pulls when it is missing locally. |
| `STRIX_RUNTIME_BACKEND` | `docker` | Sandbox backend. |
| `DOCKER_HOST` | current docker context, else the default socket | Daemon address, for example `unix:///var/run/docker.sock`. When unset, Strix uses the active docker context (Docker Desktop, OrbStack, Colima). |
| `STRIX_DOCKER_SANDBOX_NETWORK` | unset | Docker network for the sandbox. Setting it drops published ports. |
| `STRIX_DOCKER_SANDBOX_DEVICES` | unset | Comma-separated host devices mapped into the container at the same path with `rwm` access. A tunnel inside the sandbox needs `/dev/net/tun`. |
| `STRIX_SANDBOX_MEM_LIMIT` | unset | Container memory cap. Unset leaves Docker's default. |

Session Persistence

Medium
Category
Rogue Agent
Confidence
60% confidence
Finding

Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.

Content

Scanner excerpt · references/configuration.md (reported line 15)May include surrounding context.

md
The project `.env` is not read by Strix. `uv run` loads it only when you pass `--env-file` or when uv's own default `.env` discovery is left on. This skill always passes `--env-file .env` from the checkout. `--no-env-file` skips that load.

Write secrets only in `<checkout>/.env` (gitignored) or let Strix persist them to `~/.strix/cli-config.json`. Keep keys out of the scan command line, shell history pasted into chat, and git.

## Required for a hosted model