Back to skill

Security audit

wemol-cli

Security checks for vulnerabilities and agentic risk

Overview

The skill is mostly coherent for WeMol CLI use, but its broad auto-activation can lead to external lookup, local CLI installation, and account-backed job actions without a clearly narrow user trigger.

Install only if you want agents to use WeMol for scientific lookup and job workflows. Keep host command approvals enabled, require explicit confirmation before installing the CLI, submitting paid/quota-backed jobs, downloading artifacts, or canceling jobs, and verify the intended channel/account before any account-backed action.

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
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
Findings (15)

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The declared description centers on scientific knowledge retrieval and WeMol job operations, but the documented behavior also includes installing software as a prerequisite. That mismatch can cause the host or user to invoke the skill under a lower-risk mental model while it performs materially riskier actions such as downloading and executing an installer, creating an opportunity for unexpected system modification.

Content

No source excerpt is available for this finding.

Vague Triggers

High
Category
Not specified by scanner
Confidence
96% confidence
Finding

The activation condition is extremely broad: it instructs use even when the user does not mention WeMol or request computation. In a skill that can lead to shell, network, file, and installation actions, overbroad triggering raises the chance of tool invocation in unrelated conversations, causing unnecessary data exposure, remote contact, or local changes without clear user intent.

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
90% confidence
Finding

The skill exposes or implies powerful capabilities including shell, network, filesystem, and environment access, but it does not declare any tool scope or permission boundaries. In practice this weakens host-side policy enforcement and user visibility, increasing the risk of unintended command execution, installation, downloads, or local file access when the skill is activated.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

Stating that a concept question alone is sufficient to activate the skill, without boundaries, encourages unnecessary retrieval from an external platform for general scientific questions. In context, that is risky because the same skill also contains setup and execution pathways, so ambiguous concept queries may escalate into external searches or operational guidance beyond the user's expectations.

Content

No source excerpt is available for this finding.

Whitespace Padding

Medium
Category
Prompt Injection
Confidence
70% confidence
Finding

Large whitespace padding was detected (a block of blank lines or a long run of spaces). This can push injected instructions below or to the right of the visible area so a human reviewer never sees them while the agent still reads them. Manual review of the hidden content is recommended.

Content

Scanner excerpt · guides/nanobody-design.md (reported line 33)May include surrounding context.

md
The three design branches are independent once their inputs are ready.

| Method     | Intended configuration and constraints                                                                                                                                                                                                        |
| ---------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| BoltzGen   | De novo nanobody design using the prepared antigen, selected antigen chains, design count, and hotspot if specified. This workflow uses 1,000 samples.                                                                                        |
| mBER       | Nanobody design using the prepared antigen, resolved chain selection, design count, and hotspot if specified. The method may default to all chains; record the actual selected chain set for reproducibility.                                 |

Whitespace Padding

Medium
Category
Prompt Injection
Confidence
70% confidence
Finding

Large whitespace padding was detected (a block of blank lines or a long run of spaces). This can push injected instructions below or to the right of the visible area so a human reviewer never sees them while the agent still reads them. Manual review of the hidden content is recommended.

Content

Scanner excerpt · guides/nanobody-design.md (reported line 35)May include surrounding context.

md
| Method     | Intended configuration and constraints                                                                                                                                                                                                        |
| ---------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| BoltzGen   | De novo nanobody design using the prepared antigen, selected antigen chains, design count, and hotspot if specified. This workflow uses 1,000 samples.                                                                                        |
| mBER       | Nanobody design using the prepared antigen, resolved chain selection, design count, and hotspot if specified. The method may default to all chains; record the actual selected chain set for reproducibility.                                 |
| RFantibody | De novo nanobody design with a single-chain antigen. This workflow uses sequential antigen residue numbering from 1 for hotspots and a design count per template. Verify this convention and the effective template set before applying them. |

Whitespace Padding

Medium
Category
Prompt Injection
Confidence
70% confidence
Finding

Large whitespace padding was detected (a block of blank lines or a long run of spaces). This can push injected instructions below or to the right of the visible area so a human reviewer never sees them while the agent still reads them. Manual review of the hidden content is recommended.

Content

Scanner excerpt · guides/nanobody-design.md (reported line 65)May include surrounding context.

md
| Method     | Intended configuration and constraints                                                                                                                                                                                                        |
| ---------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| BoltzGen   | De novo nanobody design using the prepared antigen, selected antigen chains, design count, and hotspot if specified. This workflow uses 1,000 samples.                                                                                        |
| mBER       | Nanobody design using the prepared antigen, resolved chain selection, design count, and hotspot if specified. The method may default to all chains; record the actual selected chain set for reproducibility.                                 |
| RFantibody | De novo nanobody design with a single-chain antigen. This workflow uses sequential antigen residue numbering from 1 for hotspots and a design count per template. Verify this convention and the effective template set before applying them. |

Whitespace Padding

Medium
Category
Prompt Injection
Confidence
70% confidence
Finding

Large whitespace padding was detected (a block of blank lines or a long run of spaces). This can push injected instructions below or to the right of the visible area so a human reviewer never sees them while the agent still reads them. Manual review of the hidden content is recommended.

Content

Scanner excerpt · guides/nanobody-design.md (reported line 61)May include surrounding context.

md
Use these evaluations:

| Evaluation                                   | Input and intended settings                                                                                                                        |
| -------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------- |
| Binding Free Energy (dG)                     | Selected complex structures, with VHH chain(s) identified as the ligand. Normalize ligand chain identifiers if a shared batch setting requires it. |
| OpenDDE                                      | Selected complex sequences in batch mode; seed 39 and the antibody-antigen checkpoint `abag`.                                                      |

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The documentation instructs the agent to issue a destructive cancellation command when the user requests it, but provides no guardrails such as confirmation for ambiguous requests, identification of the exact job to cancel, or warning about irreversible interruption of running analyses. In an agent setting, this increases the risk of unintended job termination from misinterpretation, prompt injection, or stale context, especially because job cancellation affects active scientific compute workflows and can disrupt ongoing work or lose progress.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

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.

Content

Scanner excerpt · references/setup.md (reported line 67)May include surrounding context.

md
`status` checks the local runtime, not platform authentication. Inspect each channel's `access.state` in `channels`; an empty list means accounts still need configuration. Public document search/read does not require platform authentication, but CLI 1.2.0 still requires a configured channel. In a clean installation with no channels, it returns `no_channels`; follow the login setup rather than inventing a channel or claiming anonymous access works without one. Account access and computation require valid authentication.

Users manage accounts and policies in `wemol-cli tui`. New standalone channels default to no approval; existing policies survive upgrades. This does not authorize an agent to submit an unrequested computation. Platform computation can consume quota or incur charges under the user's account; public Skill installation does not include compute credits. Setup verification must not submit a job or change approval policy.

## Existing installations

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · scripts/install_cli.py (reported line 34)May include surrounding context.

python
machine = os.environ.get("PROCESSOR_ARCHITEW6432", machine).lower()
    if system == "Darwin" and machine in ("x86_64", "amd64"):
        # A translated Python process may report x86_64 on Apple Silicon.
        result = subprocess.run(["/usr/sbin/sysctl", "-in", "sysctl.proc_translated"],
                                capture_output=True, text=True, check=False)
        if result.returncode == 0 and result.stdout.strip() == "1":
            machine = "arm64"

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · scripts/install_cli.py (reported line 78)May include surrounding context.

python
def version_of(path):
    p = subprocess.run([str(path), "--version"], capture_output=True, text=True,
                       timeout=15, check=False)
    match = re.fullmatch(r"wemol-cli (\d+\.\d+\.\d+(?:[-+][\w.-]+)?)", p.stdout.strip())
    if p.returncode or not match:

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · scripts/install_cli.py (reported line 112)May include surrounding context.

python
raise RuntimeError("Release manifest must use the official HTTPS release origin")
    curl = shutil.which("curl")
    if curl:
        p = subprocess.run([curl, "--fail", "--location", "--silent", "--show-error",
                            "--proto", "=https", "--proto-redir", "=https",
                            "--connect-timeout", "15", "--max-time", "120",
                            "--max-filesize", str(asset["archive_size"]),

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The manifest describes a skill for using wemol-cli to search documentation, run analyses, and manage WeMol jobs and artifacts. This file instead performs local software installation by downloading a release archive, creating directories, and linking a new executable into the user's bin location, which is a materially different behavior than the operational capabilities described in the manifest.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Low
Category
Not specified by scanner
Confidence
92% confidence
Finding

The docstring states that 'Account data, services, PATH and shell files are untouched,' which suggests no interaction with external services. However, the implementation downloads release artifacts over HTTPS from GitHub, so the documentation understates and partially contradicts the script's actual interaction with remote services.

Content

No source excerpt is available for this finding.