Back to skill

Security audit

Siphonclaw Skill

Security checks for vulnerabilities and agentic risk

Overview

SiphonClaw appears purpose-aligned, but it asks users to run an unverified remote installer and handles broad document/image data with cloud and web-search paths that are not clearly scoped.

Review before installing. Use a safer Ollama installation method instead of piping curl to sh, inspect and pin external dependencies where possible, and avoid ingesting sensitive or regulated documents until you understand what is stored locally and what may be sent to OpenRouter, Brave Search, Telegram, AgentMail, or other configured services.

Vulnerability Patterns
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • 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
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Findings (1)

T03 · Remote Payload Retrieval and Execution

Error
Location
SKILL.md:254
Finding
Unverified Remote Installer Is Piped Directly into a Shell## Vulnerability Details **File Location**: `SKILL.md:254` **Vulnerability Type**: Remote payload retrieval and execution **Risk Level**: High **Vulnerable code:** ```bash curl -fsSL https://ollama.com/install.sh | sh ``` ### Technical Analysis The setup instructions download a mutable shell script from an external URL and execute it immediately. The command provides no version pinning, checksum verification, cryptographic signature validation, or opportunity to inspect the downloaded content before execution. HTTPS protects the connection in transit under normal conditions, but it does not establish that the remote script is immutable or safe. If the Ollama website, hosting infrastructure, DNS resolution, certificate trust path, or release process is compromised, the command could deliver arbitrary attacker-controlled shell commands. Ollama supports the skill's documented local OCR and embedding mode, but piping its installer directly into a shell is not required for that functionality. The documentation also provides a full-cloud mode that does not require Ollama. A package-manager installation or separately downloaded and verified artifact would meet the functional requirement with less risk and better adherence to least privilege. ### Attack Path 1. An attacker compromises the remote installer, its hosting infrastructure, or another trusted part of its delivery path. 2. The attacker modifies the response served from `https://ollama.com/install.sh` to include malicious shell commands. 3. A user follows the installation instructions in `SKILL.md`. 4. `curl` retrieves the current attacker-controlled response. 5. The pipe sends the response directly to `sh`, without integrity verification or review. 6. The malicious commands execute with the privileges of the user running the setup command. 7. The payload can access resources available to that user and may attempt further privilege escalation or persistence, although ...[truncated 1023 chars]
Remediation
## Remediation Suggestions 1. Remove the `curl | sh` pipeline from the setup instructions. 2. Prefer an official operating-system package manager and pin an explicit Ollama version where supported. 3. If a standalone installer is necessary: - Download it to a local file as a separate step. - Use a version-specific, immutable release URL. - Obtain the expected checksum or signature through an authenticated release channel. - Verify the artifact before execution. - Inspect the script and execute it only after successful verification. 4. Document the permissions and filesystem changes required by the installer, and advise users to avoid running the entire setup as `root`. 5. Fail closed when verification cannot be completed; do not continue with an unverified artifact. 6. For reproducible installation, pin Python dependencies and publish lock files or hash-verified requirements for the externally referenced implementation. 7. Retain the documented full-cloud mode as an alternative for users who do not need or trust the local Ollama installation path. A safer pattern is: ```bash curl -fL --output ollama-install.sh \ https://example.invalid/versioned/ollama-install.sh echo "EXPECTED_SHA256 ollama-install.sh" | sha256sum --check - less ollama-install.sh sh ollama-install.sh ``` The placeholder URL and checksum must be replaced with a versioned official release artifact and its independently published expected digest.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Tool MisuseTool Parameter Abuse, Chaining Abuse, Unsafe Defaults
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (7)

Credential Access

High
Category
Privilege Escalation
Content
pip install -r requirements.txt

# 2. Configure
cp .env.example .env
# Edit .env with your API keys (see Configuration below)

# 3. Install local models (recommended - free, unlimited ingestion)
Confidence
60% confidence
Finding
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Credential Access

High
Category
Privilege Escalation
Content
# 2. Configure
cp .env.example .env
# Edit .env with your API keys (see Configuration below)

# 3. Install local models (recommended - free, unlimited ingestion)
ollama pull qwen3-vl:latest      # 6.1 GB - OCR and vision
Confidence
60% confidence
Finding
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

External Script Fetching

High
Category
Supply Chain
Content
pip install -r requirements.txt

# 2. Install Ollama and pull local models (~10 GB total)
curl -fsSL https://ollama.com/install.sh | sh
ollama pull qwen3-vl:latest          # 6.1 GB - OCR
ollama pull bge-m3:latest             # ~1.5 GB - text embeddings
ollama pull qwen3-vl-embed:2b        # ~2 GB - visual embeddings
Confidence
97% confidence
Finding
The setup instructions recommend fetching a remote script and piping it directly to a shell, which executes unaudited code from the network with the user's privileges. If the upstream site, transport path, or published install script is compromised, users could suffer arbitrary code execution and full workstation or server compromise.

Chaining Abuse

High
Category
Tool Misuse
Content
pip install -r requirements.txt

# 2. Install Ollama and pull local models (~10 GB total)
curl -fsSL https://ollama.com/install.sh | sh
ollama pull qwen3-vl:latest          # 6.1 GB - OCR
ollama pull bge-m3:latest             # ~1.5 GB - text embeddings
ollama pull qwen3-vl-embed:2b        # ~2 GB - visual embeddings
Confidence
96% confidence
Finding
The shell pipeline chains remote content directly into command execution, eliminating any review boundary between download and execution. In practice this amplifies supply-chain risk and makes command-injection or malicious installer behavior immediately exploitable if the fetched content is altered.

Missing User Warnings

Medium
Confidence
87% confidence
Finding
The README encourages ingestion from shared drives, URLs, email, and Telegram and promotes broad indexing/searchability, but it does not warn users that sensitive documents may be transmitted to cloud APIs, persisted in local/vector stores, or exposed through external channels. In a document-intelligence system handling enterprise manuals, contracts, contact lists, and field notes, omission of data-handling and privacy cautions materially increases the risk of accidental sensitive-data exposure.

Missing User Warnings

Medium
Confidence
97% confidence
Finding
The skill states that image identification may fall back to web search when local confidence is low, but it does not give a user-facing warning that image-derived text, identifiers, or context could be shared with an external search provider. Because users may submit photos of internal equipment, labels, or error screens, this creates a meaningful risk of unintentional external disclosure.

Missing User Warnings

Medium
Confidence
95% confidence
Finding
The skill documents a hybrid/cloud mode that sends document and image-derived content to external providers, but it does not clearly warn users that potentially sensitive data may leave the local environment. In a document-intelligence pipeline, this omission can lead to accidental disclosure of proprietary, personal, or regulated content during ingestion or query-time processing.

Static analysis

No suspicious patterns detected.