Back to skill

Security audit

PPT Smith

Security checks for vulnerabilities and agentic risk

Overview

Some compatibility tools can delete previous presentation work before a replacement succeeds, creating a risk of lost files.

Use fresh, dedicated work and output directories and keep backups before using compatibility builders or rendering commands; their deletion behavior warrants review despite safer transactions in the newer template editor. Prefer fixed declarative workflows and authorize advanced Python only after reviewing the script and imports. Optional external image tools may receive selected presentation content. Scanner coverage was partial; material executable gaps were investigated statically, but runtime isolation, Office behavior and installed dependencies were not tested.

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
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • System Prompt LeakageDirect Leakage, Indirect Extraction, Tool-Based Exfiltration
Findings (83)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The supplied code is a nonfunctional visual-only adapter rather than an implementation of the declared presentation-building workflow. Although it reports itself as available, every build attempt returns failure. No undeclared resource access or unrelated behavior is present. This assessment is limited to the supplied chunk; other components could implement the broader declared functionality.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

The supplied chunk's primary function is security-boundary verification, not PPTX processing or a directly demonstrated presentation-support operation. Its network and filesystem probes are not represented in the declared description. The scope is limited: the TCP target is loopback, and the out-of-directory file is a sentinel created by the probe itself; the code does not demonstrate internet communication or access to existing user files.

Content

No source excerpt is available for this finding.

Component contract enforcement excludes modifications to existing shapes.

High
Category
Excessive Agency
Confidence
98% confidence
Finding

The author script receives the mutable slide, but subsequent contract checks inspect only shapes whose IDs were not present before execution. A script can modify or delete existing content, change existing images, or reuse existing IDs while adding one compliant shape, allowing unauthorized template or business-content changes to survive the gate. The trusted-code loading requirement limits exposure but does not enforce the advertised output contract.

Content

No source excerpt is available for this finding.

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 · engine/native_template/rendering.py (reported line 30)May include surrounding context.

python
profile = output_dir / 'profile'; profile.mkdir()
        home = output_dir / 'home'; home.mkdir()
        tmp = output_dir / 'tmp'; tmp.mkdir()
        env = dict(os.environ)
        env.update(HOME=str(home), TMPDIR=str(tmp), TMP=str(tmp), TEMP=str(tmp),
                   XDG_CONFIG_HOME=str(home / '.config'), XDG_CACHE_HOME=str(home / '.cache'),
                   XDG_DATA_HOME=str(home / '.local' / 'share'))

Unvalidated Output Injection

High
Category
Output Handling
Confidence
95% confidence
Finding

Model output is used without validation or sanitization. Unvalidated output injected into downstream contexts (SQL, shell, HTML) enables injection attacks and arbitrary code execution.

Content

Scanner excerpt · engine/pptxgenjs_builder.py (reported line 70)May include surrounding context.

python
plan_path.write_text(json.dumps(plan, ensure_ascii=False, indent=2),
                         encoding="utf-8")

    proc = subprocess.run(
        [capability["node"], str(_SCRIPT), str(plan_path), str(output.parent)],
        cwd=str(_RUNTIME_DIR), text=True, capture_output=True, check=False,
        timeout=120)

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 · packaging/bundle-exclude.txt (reported line 15)May include surrounding context.

text
cache:*.pyc

# Credentials and local environment files are always forbidden.
.env
*.pem
*.key
*.p12

Credential Access

High
Category
Privilege Escalation
Confidence
70% confidence
Finding

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

Content

Scanner excerpt · packaging/bundle-exclude.txt (reported line 20)May include surrounding context.

text
*.key
*.p12
*.pfx
credentials.json
.last-token
.netrc

Credential Access

High
Category
Privilege Escalation
Confidence
70% confidence
Finding

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

Content

Scanner excerpt · scripts/audit_bundle.py (reported line 70)May include surrounding context.

python
SECRET_ASSIGNMENT = re.compile(r"(?i)(api[_-]?key|secret|password|passwd|token|bearer|private[_-]?key)\s*[:=]\s*['\"]?([A-Za-z0-9_\-/+]{12,})")

SECRET_SHAPED_NAMES = (".env", "*.pem", "*.key", "id_rsa", "id_rsa.pub", "*.p12", "*.pfx", "credentials.json", ".last-token", ".netrc")

# A 32+ character hex or base64 run. Content digests, renderer build hashes, and
# VCS revisions are legitimately long and high-entropy, so a line is exempt when

Credential Access

High
Category
Privilege Escalation
Confidence
80% confidence
Finding

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

Content

Scanner excerpt · packaging/bundle-exclude.txt (reported line 22)May include surrounding context.

text
*.pfx
credentials.json
.last-token
.netrc

# Runtime distributions must not carry development-only material.
tests/

Credential Access

High
Category
Privilege Escalation
Confidence
80% confidence
Finding

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

Content

Scanner excerpt · scripts/audit_bundle.py (reported line 70)May include surrounding context.

python
SECRET_ASSIGNMENT = re.compile(r"(?i)(api[_-]?key|secret|password|passwd|token|bearer|private[_-]?key)\s*[:=]\s*['\"]?([A-Za-z0-9_\-/+]{12,})")

SECRET_SHAPED_NAMES = (".env", "*.pem", "*.key", "id_rsa", "id_rsa.pub", "*.p12", "*.pfx", "credentials.json", ".last-token", ".netrc")

# A 32+ character hex or base64 run. Content digests, renderer build hashes, and
# VCS revisions are legitimately long and high-entropy, so a line is exempt when

Env Variable Harvesting

High
Category
Data Exfiltration
Confidence
80% 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 · ppt_qa/slide_inspector.py (reported line 294)May include surrounding context.

python
) -> Optional[float]:
    name = obj.name.lower()
    if footnote_minimum is not None and any(
        token in name for token in SMALL_TEXT_ROLE_TOKENS
    ):
        return footnote_minimum
    return body_minimum

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 · scripts/audit_bundle.py (reported line 70)May include surrounding context.

python
SECRET_ASSIGNMENT = re.compile(r"(?i)(api[_-]?key|secret|password|passwd|token|bearer|private[_-]?key)\s*[:=]\s*['\"]?([A-Za-z0-9_\-/+]{12,})")

SECRET_SHAPED_NAMES = (".env", "*.pem", "*.key", "id_rsa", "id_rsa.pub", "*.p12", "*.pfx", "credentials.json", ".last-token", ".netrc")

# A 32+ character hex or base64 run. Content digests, renderer build hashes, and
# VCS revisions are legitimately long and high-entropy, so a line is exempt when

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 · scripts/audit_bundle.py (reported line 118)May include surrounding context.

python
stripped = stripped.split(":", 1)[1].strip()
        if stripped:
            rules.append((stripped, severity))
    return rules


def load_allowlist(path: Path) -> dict[str, set[str]]:

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 · scripts/build_public_bundle.py (reported line 37)May include surrounding context.

python
value = line.strip()
        if value and not value.startswith("#"):
            rules.append(value)
    return rules


def ignored(relative: str, rules: list[str]) -> bool:

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 · scripts/build_public_bundle.py (reported line 81)May include surrounding context.

python
"archive",
        "--fail-on-finding",
    ]
    environment = dict(os.environ)
    environment["PPTSMITH_AUDIT_TIMESTAMP"] = git("show", "-s", "--format=%cI", "HEAD")
    subprocess.run(command, cwd=destination, env=environment, check=True)

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

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 · builders/pptxgenjs_adapter.py (reported line 44)May include surrounding context.

python
errors.append("JSZIP_MODULE_NOT_FOUND")
        if node is not None and module_manifest.is_file() and jszip_manifest.is_file():
            try:
                load_check = subprocess.run(
                    [node, "--input-type=module", "-e", "await import('pptxgenjs'); await import('jszip');"],
                    cwd=self.runtime_dir,
                    text=True,

Existing deck and runtime-result files are deleted without disclosure before the replacement build succeeds (L133–L134).

Medium
Category
Output Handling
Confidence
98% confidence
Finding

The adapter unconditionally deletes the existing deck and runtime result before the replacement process succeeds. A timeout, runtime failure, or invalid result can therefore destroy the previous valid output, making this especially consequential during presentation revisions.

Content

No source excerpt is available for this finding.

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 · builders/pptxgenjs_adapter.py (reported line 142)May include surrounding context.

python
)
        command = [capability.command, str(self.runtime_dir / "build-deck.mjs"), str(input_path), str(output_dir)]
        try:
            completed = subprocess.run(
                command,
                cwd=self.runtime_dir,
                text=True,

Unbounded Output

Medium
Category
Output Handling
Confidence
60% confidence
Finding

Output size or generation rate is not bounded. Unbounded output enables denial-of-service through resource exhaustion, log flooding, or context-window stuffing.

Content

Scanner excerpt · engine/__main__.py (reported line 122)May include surrounding context.

python
c.add_argument(
        "--final-delivery",
        action="store_true",
        help="require model-authored IR, complete speaker notes, and no truncation",
    )

    rc = sub.add_parser("refine-code",

Chart verification claims to read the embedded editable workbook but reads chart XML values instead.

Medium
Category
Excessive Agency
Confidence
97% confidence
Finding

The function reads categories and series through python-pptx's chart interface, which exposes chart XML data rather than independently reading the embedded workbook. An adversarial deck can therefore pass this check with matching cached chart values but different editable workbook contents, undermining the skill's assurance that editable chart data matches the source.

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 · engine/design_scene/cli.py (reported line 349)May include surrounding context.

python
result = draft if args.full else {"status": "independent_review_required", "file": output,
                                                      "inherited_page_count": len(inherited), "inherited_pages": inherited[:8],
                                                      "pages_requiring_review": [p["page_id"] for p in draft["pages"]
                                                                                 if p["page_id"] not in inherited][:8]}
                else:
                    review = read_review(store, args.review_file)
                    operation = deliver if args.command == "deliver" else assess

Delivery documentation is always generated in Chinese, regardless of the task's selected language.

Medium
Category
Output Handling
Confidence
97% confidence
Finding

The delivery notes and their filename are hardcoded in Chinese, although tasks have a configurable language field. This imposes a language on user-facing output without checking the user's selection or documenting a justified locale restriction.

Content

No source excerpt is available for this finding.

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 · engine/design_scene/runtime.py (reported line 57)May include surrounding context.

python
"""No shell; terminate the process tree after a bounded wait."""
    options = ({"creationflags": subprocess.CREATE_NEW_PROCESS_GROUP} if os.name == "nt"
               else {"start_new_session": True})
    p = subprocess.Popen(args, cwd=cwd, env=env, stdout=subprocess.PIPE,
                         stderr=subprocess.PIPE, **options)
    try:
        out, err = p.communicate(timeout=timeout)

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 · engine/design_scene/runtime.py (reported line 65)May include surrounding context.

python
if os.name == "nt":
            killer = Path(os.environ.get("SystemRoot", "C:/Windows")) / "System32/taskkill.exe"
            try:
                subprocess.run([str(killer), "/PID", str(p.pid), "/T", "/F"],
                               stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL, timeout=10)
            except (OSError, subprocess.TimeoutExpired):
                pass