Back to plugin

Security audit

LaTeX Paper Skills

Security checks for vulnerabilities and agentic risk

Overview

The package broadly matches its academic-paper purpose, but it needs review because it can delegate workspace tasks to external AI CLIs and optionally copy Gemini OAuth/session files into a project-local directory.

Install only if you are comfortable with optional Claude/Gemini CLI delegation from your paper workspace. Avoid using --seed-user-auth in shared or version-controlled projects unless you understand that it copies Gemini OAuth/account state into a local .gemini_cli_home directory; keep that directory ignored and delete it when finished. Do not pass secrets in collaborator prompts, and review generated edits before applying them.

Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • System Prompt LeakageDirect Leakage, Indirect Extraction, Tool-Based Exfiltration
  • YARA SignaturesMalware Match, Webshell Match, Cryptominer Match
Findings (81)

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The skill claims to support empirical paper writing, but the observed behavior includes orchestrating experiments, reading matrix/config files, filtering experiment types, and writing results. In this context, the mismatch matters because experiment runners can execute substantial code and modify artifacts, which is materially more dangerous than prose editing or citation auditing.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

The skill claims to support empirical paper writing, but the observed behavior includes orchestrating experiments, reading matrix/config files, filtering experiment types, and writing results. In this context, the mismatch matters because experiment runners can execute substantial code and modify artifacts, which is materially more dangerous than prose editing or citation auditing.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The skill claims to support empirical paper writing, but the observed behavior includes orchestrating experiments, reading matrix/config files, filtering experiment types, and writing results. In this context, the mismatch matters because experiment runners can execute substantial code and modify artifacts, which is materially more dangerous than prose editing or citation auditing.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The skill claims to support empirical paper writing, but the observed behavior includes orchestrating experiments, reading matrix/config files, filtering experiment types, and writing results. In this context, the mismatch matters because experiment runners can execute substantial code and modify artifacts, which is materially more dangerous than prose editing or citation auditing.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The skill claims to support empirical paper writing, but the observed behavior includes orchestrating experiments, reading matrix/config files, filtering experiment types, and writing results. In this context, the mismatch matters because experiment runners can execute substantial code and modify artifacts, which is materially more dangerous than prose editing or citation auditing.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

The skill claims to support empirical paper writing, but the observed behavior includes orchestrating experiments, reading matrix/config files, filtering experiment types, and writing results. In this context, the mismatch matters because experiment runners can execute substantial code and modify artifacts, which is materially more dangerous than prose editing or citation auditing.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The skill claims to support empirical paper writing, but the observed behavior includes orchestrating experiments, reading matrix/config files, filtering experiment types, and writing results. In this context, the mismatch matters because experiment runners can execute substantial code and modify artifacts, which is materially more dangerous than prose editing or citation auditing.

Content

No source excerpt is available for this finding.

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · skills/latex-paper-skills/SKILL.md (reported line 34)May include surrounding context.

md
Determine the absolute directory containing this `SKILL.md`; call it

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
94% confidence
Finding

The skill launches an external AI CLI through subprocess.Popen, creating a second execution and decision-making channel not justified by the stated LaTeX paper purpose. That increases the attack surface substantially: prompts, data, and workspace context may be forwarded to another agent with different tool permissions, logging, or trust boundaries.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
96% confidence
Finding

This script exposes broad agent orchestration features—custom agents, tool allow/deny lists, MCP configuration, permission handlers, and system-prompt overrides—that are far beyond the declared purpose of a LaTeX paper-writing skill. In skill ecosystems, this kind of hidden general-purpose agent broker can be abused to alter execution context, expand access to tools, or route work through an unintended secondary agent.

Content

No source excerpt is available for this finding.

YARA rule 'agent_skill_mcp_tool_poisoning_metadata': MCP/tool metadata poisoning indicators in tool schemas or skill manifests [agent_skills]

High
Category
YARA Match
Confidence
80% confidence
Finding

YARA rule matched a hack tool or exploit indicator (offensive tools, reconnaissance, privilege escalation, or exploit frameworks).

Content

Scanner excerpt · skills/latex-paper-skills/references/skills/collaborating-with-claude/scripts/claude_bridge.py (reported line 142)May include surrounding context.

python
t
    # represent certain Unicode characters returned by Claude (e.g., "m³/s").
    # Avoid crashing on stdout/stderr encoding errors in those environments.
    try:
        sys.stdout.reconfigure(errors="backslashreplace")
    except Exception:
        pass
    try:
        sys.stderr.reconfigure(errors="backslashreplace")
    except Exception:
        pass

    parser = argparse.ArgumentParser(description="Claude Code Bridge")
    parser.add_argument("--PROMPT", required=True, help="Instruction for the task to send to claude.")
    parser.add_argument("--cd", required=True, type=Path, help="Set the workspace root for claude before executing the task.")
    session_group = parser.add_mutually_exclusive_group()
    session_group.add_argument("--SESSION_ID", default="", help="Resume the specified session of claude.")
    session_group.add_argument("--session-id", dest="session_id", default="", help="Use a specific session ID (UUID).")
    session_group.add_argument("--continue", dest="

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The MCP, agent, and tool configuration flags let callers reshape the bridge's capabilities at runtime, including loading external configs and changing permission behavior. For a LaTeX research-paper skill, this is disproportionate and dangerous because it enables hidden capability escalation and tool-context manipulation unrelated to paper drafting.

Content

No source excerpt is available for this finding.

Agent Config Directory Access

High
Category
Agent Snooping
Confidence
90% confidence
Finding

Skill reads from agent configuration directories (.claude/, .codex/, .gemini/). These directories may contain API keys, personal settings, and other credentials that the skill has no legitimate need to access.

Content

Scanner excerpt · skills/latex-paper-skills/references/skills/collaborating-with-claude/workflow.md (reported line 53)May include surrounding context.

md
Claude Code supports model aliases, so you can use `--model sonnet` / `--model opus` instead of hard-coding versioned model IDs.

- If you omit `--model`, Claude Code uses its configured default (typically from `~/.claude/settings.json`, optionally overridden by `.claude/settings.json` and `.claude/settings.local.json`).
- If you need strict reproducibility, pass a full model name via `--model <full-name>`.
- For `claim-stress-test` and `route-review-vs-empirical` templates, prefer `opus`.
- For `evidence-sufficiency-audit`, `sonnet` is sufficient.

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
92% confidence
Finding

The file acts as a bridge that executes an external Gemini CLI and streams responses, which is materially different from a narrowly scoped LaTeX authoring utility. This broadens the skill's effective capability to arbitrary delegated model interaction and external tool execution, making the mismatch between declared purpose and actual behavior a security concern because users may grant trust inappropriate to the true function.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

This function copies OAuth credentials and account state from the user's profile into a workspace-controlled directory, expanding the exposure of sensitive authentication material beyond its original protected location. In the context of a LaTeX paper-writing skill, credential replication is not necessary and increases the risk of token theft, accidental inclusion in artifacts, or reuse by other processes with access to the repo/workspace.

Content

No source excerpt is available for this finding.

YARA rule 'agent_skill_mcp_tool_poisoning_metadata': MCP/tool metadata poisoning indicators in tool schemas or skill manifests [agent_skills]

High
Category
YARA Match
Confidence
80% confidence
Finding

YARA rule matched a hack tool or exploit indicator (offensive tools, reconnaissance, privilege escalation, or exploit frameworks).

Content

Scanner excerpt · skills/latex-paper-skills/references/skills/collaborating-with-gemini/scripts/gemini_bridge.py (reported line 146)May include surrounding context.

python
emini_cli_home / ".gemini").resolve()
    dst_dir.mkdir(parents=True, exist_ok=True)

    for name in ("settings.json", "oauth_creds.json", "google_accounts.json", "state.json", "trustedFolders.json"):
        src = src_dir / name
        dst = dst_dir / name
        if src.exists() and not dst.exists():
            shutil.copy2(src, dst)


def main() -> None:
    parser = argparse.ArgumentParser(description="Gemini Bridge")
    parser.add_argument("--PROMPT", required=True, help="Instruction for the task to send to gemini.")
    parser.add_argument("--cd", required=True, type=Path, help="Set the workspace root for gemini before executing the task.")
    parser.add_argument("--sandbox", action="store_true", default=False, help="Run in sandbox mode. Defaults to `False`.")
    parser.add_argument(
        "--prefer-oauth",
        action="store_true",
        default=False,
        help="Unset API-key env vars (e.g., GEMINI_API_KEY) so Gemini CLI can use existing OAuth creds in the selec

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 · skills/latex-paper-skills/references/skills/collaborating-with-gemini/scripts/gemini_bridge.py (reported line 202)May include surrounding context.

python
"error": f"The workspace root directory `{cd.absolute().as_posix()}` does not exist. Please check the path and try again.",
        }
        print(json.dumps(result, indent=2, ensure_ascii=False))
        return

    prompt = args.PROMPT

Env Variable Harvesting

High
Category
Data Exfiltration
Confidence
60% confidence
Finding

Code enumerates, copies, or searches environment variables for secrets. Bulk environment access can collect credentials unrelated to the skill's stated purpose.

Content

Scanner excerpt · skills/latex-paper-skills/references/skills/collaborating-with-gemini/scripts/gemini_bridge.py (reported line 210)May include surrounding context.

python
# some sandboxed runners this can point to an unwritable or lock-hostile location, causing the
    # CLI to hang very early (even for `--version`). Default to a writable, repo-local home.
    # Callers can override by setting GEMINI_CLI_HOME explicitly or by passing --use-user-home.
    env = os.environ.copy()
    if not env.get("GEMINI_CLI_HOME"):
        if args.use_user_home and env.get("USERPROFILE"):
            env["GEMINI_CLI_HOME"] = str(Path(env["USERPROFILE"]).resolve())

Hidden Instructions

High
Category
Prompt Injection
Confidence
70% confidence
Finding

Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Content

Scanner excerpt · skills/latex-paper-skills/references/skills/paper-from-zero/assets/outline-contract-template.md (reported line 9)May include surrounding context.

md
|---|---------|--------|---------------|--------------|
| 1 | Introduction | Frame the problem, state the contribution, and preview the structure | 8-12 | 0-1 |
| 2 | Related Work / Background | Position the work against closest prior art | 12-20 | 0-1 |
| 3 | <!-- Method / Taxonomy / Framework --> | <!-- section-specific intent --> | 8-15 | 1-2 |
| 4 | <!-- Results / Analysis --> | <!-- section-specific intent --> | 5-10 | 2-3 |
| 5 | Discussion | Synthesize findings, state limitations, outline future work | 5-8 | 0-1 |
| 6 | Conclusion | Concise summary of contribution and takeaway | 3-5 | 0 |

Hidden Instructions

High
Category
Prompt Injection
Confidence
70% confidence
Finding

Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Content

Scanner excerpt · skills/latex-paper-skills/references/skills/paper-from-zero/assets/outline-contract-template.md (reported line 16)May include surrounding context.

md
## Totals
- Target citations: <!-- 30-80 depending on paper type -->
- Target figures/tables: <!-- 5+ -->
- Target pages: <!-- from topic brief -->

Hidden Instructions

High
Category
Prompt Injection
Confidence
70% confidence
Finding

Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Content

Scanner excerpt · skills/latex-paper-skills/references/skills/paper-from-zero/assets/topic-brief-template.md (reported line 4)May include surrounding context.

md
# Topic Brief

## Topic
<!-- One-sentence research topic -->

## Scope
<!-- What is in scope and what is explicitly out of scope -->

Hidden Instructions

High
Category
Prompt Injection
Confidence
70% confidence
Finding

Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Content

Scanner excerpt · skills/latex-paper-skills/references/skills/paper-from-zero/assets/topic-brief-template.md (reported line 13)May include surrounding context.

md
<!-- Target readers: researchers, practitioners, or both -->

## Constraints
- Venue: <!-- e.g., IEEE Access, arXiv preprint -->
- Page target: <!-- e.g., 8-10 pages -->
- Deadline: <!-- if any -->
- Special requirements: <!-- e.g., double-blind, supplementary material -->

Hidden Instructions

High
Category
Prompt Injection
Confidence
60% confidence
Finding

Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Content

Scanner excerpt · skills/latex-paper-skills/references/skills/empirical-paper-writer/assets/paper-plan-template.md (reported line 1)May include surrounding context.

md
# Routing Policy

## Route to `arxiv-paper-writer`
- Surveys, tutorials, taxonomies, benchmark syntheses

Hidden Instructions

High
Category
Prompt Injection
Confidence
60% confidence
Finding

Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Content

Scanner excerpt · skills/latex-paper-skills/references/skills/empirical-paper-writer/assets/template/references.template.bib (reported line 1)May include surrounding context.

text
# Routing Policy

## Route to `arxiv-paper-writer`
- Surveys, tutorials, taxonomies, benchmark syntheses

Hidden Instructions

High
Category
Prompt Injection
Confidence
60% confidence
Finding

Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Content

Scanner excerpt · skills/latex-paper-skills/references/skills/empirical-paper-writer/references/experiment-evidence.md (reported line 1)May include surrounding context.

md
# Routing Policy

## Route to `arxiv-paper-writer`
- Surveys, tutorials, taxonomies, benchmark syntheses

Endor dependency reachability

Checked Nov 8, 2026 · 12:39 PM UTC

Not analyzed: Endor report-only scanning is not enabled for this plugin.