Back to skill

Security audit

Advanced Human Writing & AI Humanizer

Security checks for vulnerabilities and agentic risk

Overview

ClawScan marks this skill suspicious with medium confidence, mainly because it found potentially unsafe code execution and a fixer behavior it judged risky for the skill’s writing and editing purpose.

Consider the suspicious verdict before installing. In particular, review whether the package verifier’s behavior and the fixer’s possible text deletions are acceptable for your use, and whether manuscript data could be sent to a hosted provider. The scan covered 97.4% of components and was partial, so some evidence is incomplete; the findings do not establish malicious intent.

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
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • System Prompt LeakageDirect Leakage, Indirect Extraction, Tool-Based Exfiltration
Findings (30)

Memory Manipulation

High
Category
Memory Poisoning
Confidence
80% confidence
Finding

Skill manipulates agent memory, state, or stored context. Memory corruption can alter personality, override safety rules, or cause unpredictable behavior.

Content

Scanner excerpt · README.md (reported line 174)May include surrounding context.

md
`version-key.json` locally. These commands compile prompts or aggregate recorded
ratings; they never call a model. See [editorial evaluation](docs/evaluation.md).
Real-world quality still requires independent fluent reviewers; unit tests do not
prove naturalness. Compact-review and deep-rewrite character budgets are 25,000 and
34,000 respectively (quick rewrite remains below 15,000 on the regression fixture).

## Long-Form Audit And Style Unification

Direct Prompt Extraction

High
Category
System Prompt Leakage
Confidence
85% confidence
Finding

Skill contains instructions that could directly expose system prompts, internal rules, or hidden instructions to users or external parties.

Content

Scanner excerpt · README.md (reported line 362)May include surrounding context.

human-writing-skills verify --source original.md --candidate revised.md

text

You can also run directly from the source checkout with `python -m humanwriting.cli ...`. The `build` and `humanize` commands print instruction packs that can be pasted into Codex, ChatGPT, Claude, local LLM tools, or another writing agent.

## Quick Humanize

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

The supplied script has a release-management purpose rather than implementing or supporting a specific writing or editorial operation. It reads local package metadata and validates a command-line version; it performs no text generation, rewriting, or auditing. Although it could be repository maintenance tooling, its standalone primary function is materially different from the declared skill. It does not itself publish packages or access the network.

Content

No source excerpt is available for this finding.

Anti-Refusal Statement

High
Category
Anti-Refusal
Confidence
85% confidence
Finding

Skill instructs the agent to omit warnings, disclaimers, or ethical commentary. Stripping safety caveats hides risk from the user and is a common jailbreak preamble.

Content

Scanner excerpt · docs/editorial-evaluation.md (reported line 27)May include surrounding context.

md
Use 0 for severe defects, 2 for mixed quality, and 4 for no material defect found.
Use insufficient-evidence when material cannot support a judgment; numbers do not
override this verdict. Do not judge voice without samples or factual support without
sources. Empty ballots are pending. Native-reviewed cases require two distinct fluent
reviewers and source_provenance; this is an attestation, not identity verification.
Keep results separate by language, genre, and review status, and resolve disagreements.

Direct Prompt Extraction

High
Category
System Prompt Leakage
Confidence
85% confidence
Finding

Skill contains instructions that could directly expose system prompts, internal rules, or hidden instructions to users or external parties.

Content

Scanner excerpt · humanwriting/linter.py (reported line 845)May include surrounding context.

python
paragraphs = _paragraph_spans(masked)

    def enabled(rule_id: str, category: str) -> bool:
        return rule_id not in allowed and category not in allowed

    # A single cinematic opening can be useful. Repeated fingerprints across a
    # passage are the signal that the writer is reusing a presentation recipe.

Direct Prompt Extraction

High
Category
System Prompt Leakage
Confidence
85% confidence
Finding

Skill contains instructions that could directly expose system prompts, internal rules, or hidden instructions to users or external parties.

Content

Scanner excerpt · humanwriting/review.py (reported line 49)May include surrounding context.

python
f"# Genre\n\n{skill.content}\n\n# Brief / authoritative material\n\n{brief or 'Not supplied.'}\n\n" +
        f"# Version A (data)\n\n{a}\n\n# Version B (data)\n\n{b}\n"
    )
    return prompt, mapping

Memory Manipulation

High
Category
Memory Poisoning
Confidence
85% confidence
Finding

Skill manipulates agent memory, state, or stored context. Memory corruption can alter personality, override safety rules, or cause unpredictable behavior.

Content

Scanner excerpt · skills/long-form-writing.md (reported line 73)May include surrounding context.

md
Two revision rounds per affected block are permitted; an exhausted budget requires
corrected inputs or a revised brief in a new package, not a fabricated approval.
Snapshots and prior results remain inspectable. After interruption, list and resume
the package; do not recreate completed work or overwrite state files.
Incoming pending threads must remain listed or have quoted resolution evidence.
Conflicting states from parallel branches require explicit reconciliation evidence;
never select one branch's value silently.

The manuscript-review workflow instructs users to distribute tasks through conversations or API requests without a privacy warning.

Medium
Category
Output Handling
Confidence
93% confidence
Finding

The workflow directs users to submit generated review tasks, which contain manuscript excerpts and contextual material, to new conversations or separate API requests. It does not warn that using a hosted provider discloses that material to a third party or advise checking whether the material may be shared.

Content

No source excerpt is available for this finding.

The parallel review instructions omit a privacy warning for manuscript content submitted through hosted conversations or APIs.

Medium
Category
Output Handling
Confidence
93% confidence
Finding

The instructions tell users to distribute generated manuscript-review tasks into fresh conversations or separate API requests. They do not disclose that hosted review transmits manuscript excerpts and associated context to an external provider, which matters for unpublished or confidential work.

Content

No source excerpt is available for this finding.

The staged review workflow requires model conversations or API requests without warning about disclosure of manuscript data.

Medium
Category
Output Handling
Confidence
94% confidence
Finding

The workflow instructs users to run every stage in a fresh model conversation or independent API request, potentially transmitting draft text, ledgers, and supplied sources to hosted providers. Statements elsewhere that the local compiler and MCP tools do not upload drafts do not explain the privacy implications of this subsequent submission step.

Content

No source excerpt is available for this finding.

The independent-task review workflow omits a warning about sharing manuscript material with hosted model providers.

Medium
Category
Output Handling
Confidence
93% confidence
Finding

The workflow directs users to run review tasks in new conversations or independent API calls. It does not warn that hosted execution shares the tasks' manuscript excerpts and contextual material with a third party or address confidential manuscripts.

Content

No source excerpt is available for this finding.

Skill declares no tool scope ('permissions' or 'allowed-tools') but code capabilities were detected: env, file_read, file_write, mcp, network, shell.

Medium
Category
MCP Least Privilege
Confidence
70% confidence
Finding

Without declared permissions the skill's intent is opaque and cannot be validated.

Content

No source excerpt is available for this finding.

The workflow instructs users to run manuscript-bearing prompts in model services without a privacy warning.

Medium
Category
Output Handling
Confidence
86% confidence
Finding

The generated review prompts use the supplied draft and context, and this step instructs users to run them in Chatbox or an API request. Although the document clarifies that the CLI itself does not call a model, it does not warn that using a hosted provider discloses manuscript and context content to that provider.

Content

No source excerpt is available for this finding.

The workflow directs users to submit manuscript-bearing prompts to model sessions without a privacy warning.

Medium
Category
Output Handling
Confidence
86% confidence
Finding

This step instructs users to run each generated prompt in Chatbox, an independent API request, or another model session. These prompts incorporate draft and context material, but the document does not explain that a hosted configuration sends that material to the selected provider.

Content

No source excerpt is available for this finding.

Instructions to paste drafts and continuity context into Chatbox omit a provider-disclosure warning.

Medium
Category
Output Handling
Confidence
93% confidence
Finding

The audit command accepts a manuscript and continuity ledger, then the workflow explicitly instructs users to paste its output into Chatbox. No warning explains that a remotely configured model receives this content or that unpublished writing and confidential context should be handled according to the provider's privacy and retention policies.

Content

No source excerpt is available for this finding.

The Chatbox audit workflow lacks a warning about disclosing drafts and ledger content to a hosted model.

Medium
Category
Output Handling
Confidence
93% confidence
Finding

The workflow generates audit instructions from a chapter and continuity ledger and tells users to paste the result into Chatbox. It does not warn that, when Chatbox uses a remote provider, manuscript and context content leave the local environment.

Content

No source excerpt is available for this finding.

Running manuscript chunks through API requests is instructed without a confidentiality warning.

Medium
Category
Output Handling
Confidence
85% confidence
Finding

The workflow directs users to run chunk prompts and the approved contract in fresh conversations or API requests. These inputs can contain manuscript passages, canonical context, and reference material, but there is no warning that hosted execution exposes that content to the model provider.

Content

No source excerpt is available for this finding.

The chunk-review API workflow omits a warning about transmitting manuscript and canonical context.

Medium
Category
Output Handling
Confidence
85% confidence
Finding

These instructions tell users to submit numbered chunk tasks and the confirmed baseline contract to model sessions or independent API requests. The document does not warn that hosted execution sends potentially unpublished or confidential manuscript and context material to the selected provider.

Content

No source excerpt is available for this finding.

The fixer labels context-dependent content deletion as mechanical editing.

Medium
Category
Excessive Agency
Confidence
95% confidence
Finding

The unanchored expressions delete matching phrases anywhere in the manuscript, including legitimate dialogue, quotations, and factual examples, while labeling those deletions mechanical and safe. When apply mode or an output destination is explicitly selected, this can persist silent content loss; fiction and rewrite-fidelity use cases make the mismatch particularly relevant.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
85% 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 · plugins/deepseek-harness/adapter.js (reported line 69)May include surrounding context.

js
say they are not independent agents. Do not silently reassign an already claimed task.

After each submission, re-list readiness. Fix requests permit only the bounded revisions
allowed by the returned protocol. Stop and report exhausted or blocked work; do not bypass
approval or edit package state. Later tasks stay blocked until their predecessor is approved.
Parallel work is permitted ONLY in execution_mode='contract' with approved explicit independent
block entry/exit contracts, and only for ready tasks up to max_parallel and host capacity.

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/build_skillhub_package.py (reported line 56)May include surrounding context.

python
def tracked_files(root: Path) -> list[Path]:
    result = subprocess.run(
        ["git", "ls-files", "-z"],
        cwd=root,
        check=True,

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
94% confidence
Finding

The verifier imports and executes Python modules from the supplied archive after checking paths and metadata, but those checks do not authenticate the code. If an attacker supplies the tarball, module import can execute arbitrary code with the verifier's privileges; Python isolated mode and a temporary working directory do not sandbox filesystem access, credentials, or networking.

Content

Scanner excerpt · scripts/verify_dsh_package.py (reported line 64)May include surrounding context.

python
"assert plan_long_form_writing('outline.md', 'writing', style='fiction', target_length=60000)['chunk_count'] == 12; "
                "assert [t['task_id'] for t in list_writing_tasks('writing')['tasks'] if t['ready']] == ['outline-check']"
            )
            subprocess.run([sys.executable, "-I", "-X", "utf8", "-c", probe, str(Path(directory) / "package")],
                           cwd=directory, check=True, timeout=30)
    print(f"DSH package verified: {Path(path).name}")

Dynamic attribute access via getattr()

Low
Category
Dangerous Code Execution
Confidence
50% confidence
Finding

Dynamic getattr() with a non-literal attribute name can access arbitrary object attributes, potentially bypassing access controls.

Content

Scanner excerpt · humanwriting/config.py (reported line 90)May include surrounding context.

python
if "allow" in values:
        for attribute, option in (("allow", "--allow"), ("lint_allow", "--lint-allow")):
            if hasattr(args, attribute) and not supplied(option):
                current = list(getattr(args, attribute) or [])
                setattr(args, attribute, list(dict.fromkeys([*values["allow"], *current])))
    return config

Dynamic attribute access via getattr()

Low
Category
Dangerous Code Execution
Confidence
50% confidence
Finding

Dynamic getattr() with a non-literal attribute name can access arbitrary object attributes, potentially bypassing access controls.

Content

Scanner excerpt · humanwriting/mcp_server.py (reported line 495)May include surrounding context.

python
from . import writing_queue

    try:
        return _text_result(getattr(writing_queue, name)(**values))
    except (KeyError, TypeError) as exc:
        raise ValueError(f"Invalid writing task data: {exc}") from exc

Unverifiable Dependency: setuptools has 11 known advisory(ies) (CVE-2013-1633 (Setuptools vulnerable to Man-in-the-middle attacks); CVE-2025-47273 (setuptools has a path traversal vulnerability in PackageIndex.download that lead); CVE-2024-6345 (setuptools vulnerable to Command Injection via package URL) +8 more), but the manifest does not pin a version, so it is unknown whether the installed release is affected

Low
Category
Supply Chain
Confidence
40% confidence
Finding

OSV returned advisories for this package, but its resolved version is unknown. A package-name lookup does not establish whether the installed release is affected by those advisories.

Content

No source excerpt is available for this finding.