Back to skill

Security audit

video-add-b-roll

Security checks for vulnerabilities and agentic risk

Overview

This skill is a disclosed video-editing workflow for adding B-roll, with expected local media processing, Pexels access, and project-file writes.

Install/use this only for projects where you are comfortable letting the skill modify local video project files, create review artifacts, run ffmpeg, and contact Pexels using your API key. Keep the key in the local .env or environment, review generated pages/receipts before applying them, and run it only against the intended project root.

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
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • Memory PoisoningPersistent Context Injection, Context Window Stuffing, Memory Manipulation
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
Findings (102)

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

Beyond cutaway planning, the skill includes speaker-inset composition, preview generation, media decoding/rendering, and review pipeline logic. This is a meaningful expansion of capability relative to the declared purpose, and it can hide that the skill is effectively a media-processing and project-modification workflow with substantial local execution authority.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

Beyond cutaway planning, the skill includes speaker-inset composition, preview generation, media decoding/rendering, and review pipeline logic. This is a meaningful expansion of capability relative to the declared purpose, and it can hide that the skill is effectively a media-processing and project-modification workflow with substantial local execution authority.

Content

No source excerpt is available for this finding.

Credential Access

High
Category
Privilege Escalation
Confidence
91% confidence
Finding

The skill reads a Pexels API key from a local .env file and loads it into the process environment. Although it warns against printing the key, direct secret ingestion from repository-relative files expands credential exposure to the whole skill runtime and any subprocesses it launches, which is risky in a shell-heavy workflow that also performs network actions.

Content

Scanner excerpt · SKILL.md (reported line 43)May include surrounding context.

powershell
$RepoRoot = (Resolve-Path '.').Path
$ProjectRoot = (Resolve-Path 'path/to/video-project').Path
$BrollEnv = Join-Path $RepoRoot 'skills/video-add-b-roll/.env'
if ([string]::IsNullOrWhiteSpace($env:PEXELS_API_KEY) -and (Test-Path -LiteralPath $BrollEnv)) {
  $PexelsKeyLine = Get-Content -LiteralPath $BrollEnv | Where-Object { $_ -match '^\s*PEXELS_API_KEY\s*=' } | Select-Object -Last 1
  if ($PexelsKeyLine) { $env:PEXELS_API_KEY = ($PexelsKeyLine -split '=', 2)[1].Trim().Trim('"').Trim("'") }

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 · scripts/build_review_page.py (reported line 41)May include surrounding context.

python
def _extract_frame(video, time_s, output):
    subprocess.run(["ffmpeg", "-y", "-loglevel", "error", "-ss", f"{time_s:.6f}", "-i", str(video), "-frames:v", "1", "-vf", "scale=960:-2", "-q:v", "2", str(output)], check=True, capture_output=True)


def _probe_video(video):

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 · scripts/normalize_broll.py (reported line 462)May include surrounding context.

python
_check_probe(probe, width, height, num, den, _shot_duration(shot, timeline))
        if record.get("probe") != probe:
            raise ValueError("normalized probe is stale")
        subprocess.run([
            "ffmpeg", "-v", "error", "-i", str(output),
            "-map", "0:v:0", "-f", "null", "-",
        ], check=True, capture_output=True)

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 · scripts/normalize_broll.py (reported line 530)May include surrounding context.

python
_check_probe(probe, width, height, num, den, _shot_duration(shot, timeline))
        if record.get("probe") != probe:
            raise ValueError("normalized probe is stale")
        subprocess.run(
            ["ffmpeg", "-v", "error", "-i", str(output),
             "-map", "0:v:0", "-f", "null", "-"],
            check=True, capture_output=True,

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 · scripts/normalize_broll.py (reported line 590)May include surrounding context.

python
_check_probe(probe, width, height, num, den, _shot_duration(shot, timeline))
        if record.get("probe") != probe:
            raise ValueError("normalized probe is stale")
        subprocess.run(
            ["ffmpeg", "-v", "error", "-i", str(output),
             "-map", "0:v:0", "-f", "null", "-"],
            check=True, capture_output=True,

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 · scripts/normalize_broll.py (reported line 620)May include surrounding context.

python
_check_probe(probe, width, height, num, den, _shot_duration(shot, timeline))
    if record.get("probe") != probe:
        raise ValueError("normalized probe is stale")
    subprocess.run(["ffmpeg", "-v", "error", "-i", str(output), "-map", "0:v:0", "-f", "null", "-"], check=True, capture_output=True)


def _speaker_documents(plan, root):

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 · scripts/speaker_inset.py (reported line 258)May include surrounding context.

python
def _extract_frame(video, time_s, output):
    output = Path(output)
    output.parent.mkdir(parents=True, exist_ok=True)
    subprocess.run([
        "ffmpeg", "-v", "error", "-y", "-i", str(video),
        "-ss", f"{float(time_s):.9f}", "-frames:v", "1",
        "-vf", "scale=960:-2", str(output),

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 · tests/test_broll.py (reported line 5383)May include surrounding context.

python
plan, shot, self.timeline, self.root, output, None,
        )

        probe = subprocess.run([
            "ffprobe", "-v", "error", "-count_frames", "-select_streams", "v:0",
            "-show_entries", "stream=nb_read_frames", "-of", "json", str(output),
        ], check=True, capture_output=True, text=True)

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 · tests/test_broll.py (reported line 6760)May include surrounding context.

python
plan, shot, self.timeline, self.root, output, None,
        )

        probe = subprocess.run([
            "ffprobe", "-v", "error", "-count_frames", "-select_streams", "v:0",
            "-show_entries", "stream=nb_read_frames", "-of", "json", str(output),
        ], check=True, capture_output=True, text=True)

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 · tests/test_broll.py (reported line 6040)May include surrounding context.

python
self.assertFalse(result["probe"]["has_audio"])
        self.assertEqual(base_record["probe"]["fps"], result["probe"]["fps"])
        still = self.root / "delivery-speaker.png"
        subprocess.run([
            "ffmpeg", "-y", "-loglevel", "error", "-ss", "0.5",
            "-i", str(self.output), "-frames:v", "1", str(still),
        ], check=True, capture_output=True)

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 · tests/test_broll.py (reported line 6977)May include surrounding context.

python
"program_duration_s": boundary,
                "playback_rate": 1.0,
            })
        subprocess.run([
            "ffmpeg", "-y", "-loglevel", "error",
            "-i", str(normalize_broll._segment_output(self.output, 1)),
            "-i", str(normalize_broll._segment_output(self.output, 2)),

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 · tests/test_broll.py (reported line 7155)May include surrounding context.

python
self.assertEqual(self.output, record["path"])
        first, last = self.root / "first.png", self.root / "last.png"
        for time_s, frame in ((0, first), (0.8, last)):
            subprocess.run(["ffmpeg", "-y", "-loglevel", "error", "-ss", str(time_s), "-i", str(self.output), "-frames:v", "1", str(frame)], check=True, capture_output=True)
        with Image.open(first) as first_image, Image.open(last) as last_image:
            self.assertNotEqual(first_image.tobytes(), last_image.tobytes())

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 · tests/test_broll.py (reported line 7758)May include surrounding context.

python
def test_checker_accepts_duration_short_by_one_frame_and_extracts_last_still(self):
        video, normalized = self._normalized_for_check()
        subprocess.run([
            "ffmpeg", "-y", "-loglevel", "error",
            "-f", "lavfi", "-i", "testsrc2=size=96x54:rate=30000/1001",
            "-frames:v", "29", "-an", "-c:v", "libx264", "-pix_fmt", "yuv420p", str(self.output),

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 · tests/test_normalize_broll_frames.py (reported line 304)May include surrounding context.

python
normalize_broll.normalize_shot(candidate, shot, timeline, output)

            probe = subprocess.run([
                "ffprobe", "-v", "error", "-count_frames", "-select_streams", "v:0",
                "-show_entries", "stream=nb_read_frames", "-of", "json", str(output),
            ], check=True, capture_output=True, text=True)

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
93% confidence
Finding

The skill exercises broad capabilities including shell execution, filesystem reads/writes, environment access, and network use, but it declares no explicit tool scope or permission boundaries. That creates an ambient-authority problem: a caller or agent may execute powerful operations the user did not clearly authorize, increasing the blast radius of mistakes or prompt-injection-driven actions.

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 · SKILL.md (reported line 509)May include surrounding context.

md
composite-only revision creates a new composite UUID page while preserving
the approved `selection_sha256`; it cannot silently change B-roll. After composite approval,
derive candidate decisions from the approved selection and transition `composite_pending` shots
without asking for the same content decision again.

Before accepting `ambiguous`, request dense supplemental evidence inside that subshot and rebuild
the analysis hash. At clearance, inspect entry, middle, exit, and a reasoned motion-risk frame in

Context Window Stuffing

Medium
Category
Memory Poisoning
Confidence
80% confidence
Finding

Skill attempts to fill the context window with filler content, displacing legitimate instructions and safety constraints. This can degrade agent performance or bypass safety boundaries.

Content

Scanner excerpt · examples/example-broll-plan.json (reported line 75)May include surrounding context.

json
"consumed": true,
    "path": "b-roll/broll-selection.json",
    "sha256": "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa",
    "review_page_sha256": "abababababababababababababababababababababababababababababababab",
    "review_id": "123e4567-e89b-12d3-a456-426614174000",
    "mode": "human",
    "actor": "Actual user",

Context Window Stuffing

Medium
Category
Memory Poisoning
Confidence
80% confidence
Finding

Skill attempts to fill the context window with filler content, displacing legitimate instructions and safety constraints. This can degrade agent performance or bypass safety boundaries.

Content

Scanner excerpt · examples/example-broll-plan.json (reported line 164)May include surrounding context.

json
"id": "example-owned-factory-detail-video",
          "media_type": "video",
          "cache_path": "cache/b-roll/candidates/example-owned-factory-detail-video.mp4",
          "sha256": "6767676767676767676767676767676767676767676767676767676767676767",
          "bytes": 1123456,
          "duration_s": 4.0,
          "probe": {

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
80% 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 · SKILL.md (reported line 401)May include surrounding context.

md
`submission_intent: request_revision`; Empty Request changes is invalid. Reject it in
  `apply_review()` before any plan or receipt write, then use the existing validation,
  revision/rebuild, and new immutable page flow. The rebuilt proposal carries the first route
  binding forward; do not ask the user to choose ordinary or speaker-inset again. The
  `presentation-decision.json` receipt remains unchanged, and its candidate-manifest and
  review-video bindings remain mandatory. Never convert revision notes into human authority.
- Candidate selection and speaker composite approval remain separate. The first page approves B-roll

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
80% 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 · reference/broll-rules.md (reported line 143)May include surrounding context.

md
`submission_intent: request_revision`; Empty Request changes is invalid. Reject it in
  `apply_review()` before any plan or receipt write, then use the existing validation,
  revision/rebuild, and new immutable page flow. The rebuilt proposal carries the first route
  binding forward; do not ask the user to choose ordinary or speaker-inset again. The
  `presentation-decision.json` receipt remains unchanged, and its candidate-manifest and
  review-video bindings remain mandatory. Never convert revision notes into human authority.
- Candidate selection and speaker composite approval remain separate. The first page approves B-roll

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
80% 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 · tests/test_broll.py (reported line 445)May include surrounding context.

python
`submission_intent: request_revision`; Empty Request changes is invalid. Reject it in
  `apply_review()` before any plan or receipt write, then use the existing validation,
  revision/rebuild, and new immutable page flow. The rebuilt proposal carries the first route
  binding forward; do not ask the user to choose ordinary or speaker-inset again. The
  `presentation-decision.json` receipt remains unchanged, and its candidate-manifest and
  review-video bindings remain mandatory. Never convert revision notes into human authority.
- Candidate selection and speaker composite approval remain separate. The first page approves B-roll

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
80% 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 · reference/broll-rules.md (reported line 178)May include surrounding context.

md
`rationale_source: "agent_chat_explicit_action"`. The plan binds that receipt by SHA-256. This
  chat decision selects a route only; it cannot impersonate a `review_ui_explicit_action` or replace
  either webpage approval. A candidate revision may change the unapproved proposal's timing and
  segment defaults while preserving the first route binding. Do not ask the user to choose ordinary
  or speaker-inset again. The `presentation-decision.json` receipt remains unchanged. A changed
  candidate manifest or review video still requires a new chat choice.
- An `ordinary` choice removes `speaker_inset_style` and uses the existing one-page workflow. A

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_review_page.py (reported line 41)May include surrounding context.

python
def _extract_frame(video, time_s, output):
    subprocess.run(["ffmpeg", "-y", "-loglevel", "error", "-ss", f"{time_s:.6f}", "-i", str(video), "-frames:v", "1", "-vf", "scale=960:-2", "-q:v", "2", str(output)], check=True, capture_output=True)


def _probe_video(video):