Back to skill

Security audit

video-add-captions

Security checks for vulnerabilities and agentic risk

Overview

This skill is a disclosed local video-caption workflow with review gates and project-scoped file generation, though it relies on external media tools and an on-demand HyperFrames package.

Install/use this only for projects where local media processing is expected. Review the generated HTML evidence before approval, keep the working directory project-scoped, and consider pinning HyperFrames instead of relying on ambient npx resolution.

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
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Memory PoisoningPersistent Context Injection, Context Window Stuffing, Memory Manipulation
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
Findings (49)

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The declared description is for a specialized caption-rendering skill in an Open Recut workflow. The actual code chunk is a third-party front-end animation library (GSAP), not application logic for caption generation or rendering. It does not read transcripts, map timing data, process timeline.json, render image sequences, or register overlays. Its primary purpose is materially different: generic browser animation support. While such a library could be a supporting dependency in a UI, this code chunk by itself does not implement the declared functionality, so the description does not accurately represent the supplied code.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
89% confidence
Finding

The code is caption-planning/build tooling, not a full end-to-end caption rendering and registration skill as described. It does align with part of the description: it can map transcript data through a timeline and create word-timed caption plans with renderer_recipe metadata referencing HyperFrames assets. However, the substantial claimed behaviors—rendering a local transparent PNG sequence, reviewing maintained style on source-backed pixels, and registering the result as an overlay contribution—are absent. Instead, the script mainly groups words into cues, writes JSON/SRT outputs, and validates expressive caption plans. That is a material description/behavior mismatch.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

The description claims an operational caption-rendering skill: map transcript through timeline.json, review style on pixels, render a transparent PNG sequence, and register it as an overlay for delivery. The supplied code does none of that. It is a metadata tool centered on approved B-roll composite context: it reads project.json and B-roll-related artifacts, splits/aligs caption cues to visual boundaries, computes deterministic placement metadata, validates stale bindings via SHA-256, and attaches a spatial_context reference into the plan JSON. While this supports caption placement, it neither renders images nor produces overlay assets nor integrates with delivery rendering. Therefore the declared description materially overstates and misrepresents the actual behavior.

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 · SKILL.md (reported line 445)May include surrounding context.

md
The builder writes `captions-review.html` as the one authoritative page. Standard

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 595)May include surrounding context.

md
The builder writes `captions-review.html` as the one authoritative page. Standard

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 718)May include surrounding context.

md
`generate_caption_project.mjs` accepts that array with the interaction receipt.

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 · scripts/check_caption_interaction.mjs (reported line 1438)May include surrounding context.

js
check("bound v1 state cannot omit maintained style definitions", () => {
    const fixture = preparePreview("bound-missing-style-definitions");
    const state = readState(fixture.statePath);
    delete state.styleDefinitions;
    writeFileSync(fixture.statePath, JSON.stringify(state), "utf8");
    assert.throws(() => interactionState.validateGenerationInteraction({
      statePath: fixture.statePath,

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
93% confidence
Finding

The skill clearly instructs file reads, file writes, and shell execution across the repository and project directories, but it declares no explicit tool scope or permission boundaries. In an agent environment, that omission increases the blast radius of prompt-injected or mistaken execution because the runtime cannot enforce least privilege for this skill.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
91% confidence
Finding

The workflow invokes npx hyperframes, which may fetch and execute whatever package/version is currently resolved at runtime. This creates a supply-chain execution path where a compromised upstream package, typo-squatted dependency, or unexpected version change could run arbitrary code on the host during media processing.

Content

No source excerpt is available for this finding.

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 310)May include surrounding context.

--project-root $ProjectRoot --plan $Plan --context $SpatialContext

text

Do not ask the user to choose a position cue by cue. The user reviews the existing
HTML evidence and corrects only a small number of anomalous beats. Starting the
interaction after planning keeps the existing caption-plan hash binding authoritative.

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 349)May include surrounding context.

--project-root $ProjectRoot --plan $Plan --context $SpatialContext

text

Do not ask the user to choose a position cue by cue. The user reviews the existing
HTML evidence and corrects only a small number of anomalous beats. Starting the
interaction after planning keeps the existing caption-plan hash binding authoritative.

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 315)May include surrounding context.

interaction after planning keeps the existing caption-plan hash binding authoritative.

Start one decision mode. For human mode, publish the maintained offline gallery into the project review directory without asking the user to locate a file:

powershell
$StartOutput = node "$SkillRoot\scripts\caption_interaction.mjs" start `

Rp1

Medium
Category
MCP Rug Pull
Confidence
88% confidence
Finding

The script resolves npx from the current environment and runs hyperframes without any visible version pinning or trusted binary path. In a build/review pipeline, this can execute an unexpected or malicious package version from the local/npm registry context, leading to supply-chain compromise and arbitrary code execution under the privileges of the review process.

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 · scripts/build_caption_review.py (reported line 581)May include surrounding context.

python
raise RuntimeError("npx is required to capture HyperFrames preview snapshots")
    output_dir.mkdir(parents=True, exist_ok=True)
    times = ",".join(f"{item['program_s']:.12f}" for item in samples)
    subprocess.run([
        executable, "hyperframes", "snapshot", str(Path(project).resolve()),
        "--at", times, "--no-end", "--timeout", "60000", "--describe", "false",
        "--output", str(output_dir.resolve()),

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_caption_review.py (reported line 858)May include surrounding context.

python
sample, spatial_context, project_root, source, source_s, source_hash,
            )
            source_frame = source_frames / f"source-{sample['file_stem']}.png"
            subprocess.run([
                "ffmpeg", "-hide_banner", "-loglevel", "error", "-y",
                "-ss", f"{background['seek_s']:.6f}", "-i", background["path"],
                "-frames:v", "1", str(source_frame),

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/check_caption_review.py (reported line 721)May include surrounding context.

python
overlay.save(snapshots / f"frame-{index:03d}.png")
        state = root / "interaction.json"
        interaction = SCRIPT_DIR / "caption_interaction.mjs"
        subprocess.run([
            "node", str(interaction), "start", "--state", str(state.resolve()), "--source", str(source.resolve()),
            "--captions", str(plan.resolve()), "--spatial-context", str(context_path.resolve()),
            "--review-dir", str(root / "style-review"), "--decision-mode", "agent",

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/check_caption_review.py (reported line 727)May include surrounding context.

python
"--review-dir", str(root / "style-review"), "--decision-mode", "agent",
            "--delegation-note", "Synthetic review binding regression.", "--no-open", "true",
        ], check=True, capture_output=True, encoding="utf-8", errors="replace")
        subprocess.run([
            "node", str(interaction), "agent-select", "--state", str(state.resolve()), "--choice", "clean",
            "--rationale", "Use the stable clean fixture.",
        ], check=True, capture_output=True, encoding="utf-8", errors="replace")

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/check_caption_review.py (reported line 765)May include surrounding context.

python
"spatialContext": {"sha256": file_sha256(context_path)},
        })
        evidence_paths = [str((out / sample["preview"]).resolve()) for sample in evidence["review_samples"]]
        result = subprocess.run([
            "node", str(interaction), "preview-ready", "--state", str(state.resolve()),
            "--project-meta", str(project_meta), "--evidence", ",".join(evidence_paths),
            "--evidence-document", str((out / "captions-evidence.json").resolve()),

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/check_caption_spatial_context.py (reported line 362)May include surrounding context.

python
state = run / "interaction.json"
    project = run / "project"
    run.mkdir(parents=True)
    subprocess.run([
        "node", str(scripts / "caption_interaction.mjs"), "start",
        "--state", str(state), "--source", str(source), "--captions", str(plan_path),
        "--spatial-context", str(context_path), "--decision-mode", "agent",

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/check_caption_spatial_context.py (reported line 368)May include surrounding context.

python
"--spatial-context", str(context_path), "--decision-mode", "agent",
        "--delegation-note", "Synthetic spatial renderer regression.", "--no-open", "true",
    ], check=True, capture_output=True, encoding="utf-8", errors="replace")
    subprocess.run([
        "node", str(scripts / "caption_interaction.mjs"), "agent-select",
        "--state", str(state), "--choice", "clean",
        "--rationale", "Use the stable clean fixture.",

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/check_caption_spatial_context.py (reported line 373)May include surrounding context.

python
"--state", str(state), "--choice", "clean",
        "--rationale", "Use the stable clean fixture.",
    ], check=True, capture_output=True, encoding="utf-8", errors="replace")
    subprocess.run([
        "node", str(scripts / "generate_caption_project.mjs"),
        "--video", str(source), "--captions", str(plan_path), "--out", str(project),
        "--interaction-state", str(state), "--spatial-context", str(context_path),

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/check_caption_spatial_context.py (reported line 389)May include surrounding context.

python
root = Path(temporary)
        _fixture(root)
        source = root / "source.mp4"
        subprocess.run([
            "ffmpeg", "-hide_banner", "-loglevel", "error", "-y",
            "-f", "lavfi", "-i", "color=c=0x243447:s=320x180:r=30:d=12",
            "-an", "-c:v", "libx264", "-pix_fmt", "yuv420p", str(source),

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/check_caption_spatial_context.py (reported line 471)May include surrounding context.

python
context_bytes = standard_context_path.read_bytes()
        standard_context_path.write_bytes(context_bytes + b"\n")
        stale = subprocess.run([
            "node", str(Path(__file__).resolve().parent / "caption_interaction.mjs"),
            "status", "--state", str(standard_state),
        ], check=False, capture_output=True, encoding="utf-8", errors="replace")

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/check_caption_spatial_context.py (reported line 484)May include surrounding context.

python
context_bytes = standard_context_path.read_bytes()
        standard_context_path.write_bytes(context_bytes + b"\n")
        stale = subprocess.run([
            "node", str(Path(__file__).resolve().parent / "caption_interaction.mjs"),
            "status", "--state", str(standard_state),
        ], check=False, capture_output=True, encoding="utf-8", errors="replace")

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/check_project_protocol.py (reported line 327)May include surrounding context.

python
invalid = copy.deepcopy(complete)
        invalid["presentation"]["layout_beats"][0]["variant"] = "top-statement"
        invalid_path.write_text(json.dumps(invalid), encoding="utf-8")
        result = subprocess.run(
            [sys.executable, str(Path(build_captions.__file__)), "--validate-plan", str(invalid_path)],
            check=False, capture_output=True, encoding="utf-8", errors="replace",
        )