Back to skill

Security audit

550W AI Subtitle & Watermark Remover

Security checks for vulnerabilities and agentic risk

Overview

The skill mostly matches its media-removal purpose, but it exposes a destructive task-deletion action without an enforceable confirmation step.

Review this before installing if you use valuable media tasks in the 550W account. Selected media or public links are sent to the 550W API, credentials may be stored locally, and deleting a subtitle task can clean up related remote media without refunding credits. Use connector or environment secret storage when possible, and only allow deletion through a host workflow that asks for explicit confirmation.

Vulnerability Patterns
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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
Findings (1)

T09 · Insecure Skill Coding Practices

Warning
Location
dist/actions/delete-task.js:8
Finding

Destructive task deletion lacks an enforced confirmation gate

Content
View full analysis

Vulnerability Details

File Location: dist/actions/delete-task.js:8-16
Vulnerability Type: Missing authorization confirmation for a destructive operation
Risk Level: Medium

The Skill documentation requires immediate user confirmation before deleting a subtitle task (SKILL.md:38), but the executable deletion path does not require or validate proof of that confirmation.

Complete Code Snippet

javascript
async function deleteTask(params, client) {
    const validationError = (0, validator_1.validate)("deleteTask", params);
    if (validationError)
        return validationError;
    const requestParams = { taskId: params.taskId.trim() };
    if (params.subUserId != null && String(params.subUserId).trim() !== "") {
        requestParams.subUserId = String(params.subUserId).trim();
    }
    const response = await client.post("/open/deleteTask", requestParams, types_1.TIMEOUT_CONFIG.query);
    return response.code === types_1.ErrorCode.SUCCESS
        ? { ...response, message: (0, i18n_1.localize)(params.locale, "任务已删除,相关媒体将异步清理;已扣积分不会退还", "Task deleted. Related media will be cleaned up asynchronously; charged credits are not refunded.") }
        : (0, error_handler_1.mapApiError)(response, params.locale, params.region);
}

The MCP exposure in dist/mcp-server.js:94-103 accepts only a task ID, optional sub-user scope, and locale before directly dispatching deletion:

javascript
server.registerTool("delete_subtitle_task", {
    description: description("删除本人一项去字幕任务,并安排清理上传及结果媒体;已扣积分不退还。", "Permanently hide one owned subtitle-removal task and schedule its uploaded and result media for cleanup. Charged credits are not refunded."),
    inputSchema: {
        taskId: zod_1.z.string().min(1).max(64).describe(description("去字幕任务 ID。", "Task ID returned by subtitle removal.")),
        subUserId: zod_1.z.string().max(128).optional().describe(description("提交时使用的可选子用户标识。", "Option
...[truncated 2617 chars]
Remediation
View remediation

Remediation Suggestions

  1. Add a mandatory confirmation field, such as confirmed: z.literal(true), to every deletion entry point.
  2. Validate confirmation again inside deleteTask so direct dispatcher or CLI invocation cannot bypass the MCP schema.
  3. Prefer a two-step flow:
    • First return the exact task ID, affected media, cleanup consequences, and lack of credit refund.
    • After explicit user approval, issue a short-lived, single-use confirmation token bound to the account, task ID, sub-user ID, and action.
    • Require and consume that token before calling /open/deleteTask.
  4. Reject stale, reused, missing, or mismatched confirmation tokens.
  5. Apply the same gate to the MCP server, standard-input CLI, and any exported programmatic API.
  6. Keep destructiveHint: true, but treat it only as metadata rather than as an authorization control.
  7. Add tests proving that deletion requests without valid confirmation never reach ApiClient.post.
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
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • System Prompt LeakageDirect Leakage, Indirect Extraction, Tool-Based Exfiltration
Findings (56)

Tp4

High
Category
MCP Tool Poisoning
Confidence
93% confidence
Finding

The declared description focuses on media watermark/subtitle/text removal and checking task progress/credits. The supplied code chunk exposes a deleteTask action, which is a distinct task-management capability not disclosed in the description. Because deleting tasks is an undeclared capability and the actual code shown does not implement the stated primary processing behavior, this is a description-behavior mismatch based on the provided chunk.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

This code chunk's primary function is media upload, not media editing or watermark/subtitle removal as declared. While uploading could be a supporting step in a larger workflow, the declared description does not mention upload capability, and this chunk does not implement the stated core behaviors of removing subtitles/watermarks or checking tasks/credits. Therefore the behavior shown is materially different from the declared purpose.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

Yes, this is a mismatch. The declared description promises substantial media-processing functionality and API-based task/credit management, but the provided code chunk contains only an empty CLI declaration file with no operational logic. Based on this code alone, the actual behavior does not implement or evidence the described capabilities.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The declared purpose centers on watermark/subtitle/text removal and related API task/credit handling. The supplied code chunk does not implement any media processing, share-link handling, task polling, or credit checking. Instead, it manages credentials and configuration state. While credential management could be a supporting utility for such a skill, this chunk by itself does not accurately represent the declared end-user functionality and reflects a materially different primary behavior in isolation.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

This code chunk appears to be an internationalization and service-region utility type declaration file, not code for media watermark/subtitle removal or task/credit management. While helper utilities can be supporting details, the supplied chunk contains only those helpers and none of the core declared behavior. Therefore, the actual behavior represented by this chunk does not accurately match the declared purpose.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The declared description promises substantial media-editing and API-driven functionality, but the provided code chunk contains no such behavior at all. It does not process videos or images, remove watermarks or subtitles, call any external API, or check tasks/credits. Because the actual code is essentially a no-op declaration stub, it does not accurately represent the declared purpose.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

The declared purpose centers on media watermark/subtitle/text removal through the 550W Open API and task/credit checking. This code does none of that. Its primary function is to pre-validate remote video URLs for public accessibility and inspect them with ffprobe to obtain metadata. While such probing could be a supporting step in a media-processing workflow, the chunk as provided is materially different from the declared user-facing capability and lacks any implementation of the advertised API operations, watermark removal, subtitle removal, image text removal, task lookup, or credit checks.

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 · dist/550w-mcp.cjs (reported line 736)May include surrounding context.

js
get error() {
            if (this._error)
              return this._error;
            const error = new ZodError_js_1.ZodError(ctx.common.issues);
            this._error = error;
            return this._error;
          }
        };
      }
    };
    function processCreateParams(params) {
      if (!params)
        return {};
      const { errorMap, invalid_type_error, required_error, description: description2 } = params;
      if (errorMap && (invalid_type_error || required_error)) {
        throw new Error(`Can't use "invalid_type_error" or "required_error" in conjunction with custom error map.`);
      }
      if (errorMap)
        return { errorMap, description: description2 };
      const customMap = (iss, ctx) => {
        const { message } = params;
        if (iss.code === "invalid_enum_value") {
          return { message: message ?? ctx.defaultError };
        }
        if (typeof ctx.data === "undefined") {
          return { message: message ?? required_error ??

YARA rule 'exploit_framework': Exploit framework components and payloads [hacktools]

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 · dist/550w-mcp.cjs (reported line 24671)May include surrounding context.

js
...util.normalizeParams(params)
      });
    }
    // @__NO_SIDE_EFFECTS__
    function keyof(schema) {
      const shape = schema._zod.def.shape;
      return /* @__PURE__ */ _enum(Object.keys(shape));
    }
    exports2.ZodMiniObject = core.$constructor("ZodMiniObject", (inst, def) => {
      core.$ZodObject.init(inst, def);
      exports2.ZodMiniType.init(inst, def);
      util.installLazyProp(inst, "shape", (self) => self._zod.def.shape, false);
    });
    // @__NO_SIDE_EFFECTS__
    function object(shape, params) {
      const def = {
        type: "object",
        shape: shape ?? {},
        ...util.normalizeParams(params)
      };
      return new exports2.ZodMiniObject(def);
    }
    // @__NO_SIDE_EFFECTS__
    function strictObject(shape, params) {
      return new exports2.ZodMiniObject({
        type: "object",
        shape,
        catchall: /* @__PURE__ */ never(),
        ...util.normalizeParams(params)
      });
    }
    // @__NO_SIDE_EFFECTS__
    function loos

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 · dist/550w-mcp.cjs (reported line 35892)May include surrounding context.

js
function validateAsync() {
        const ruleErrs = gen.let("ruleErrs", null);
        gen.try(() => assignValid((0, codegen_1._)`await `), (e) => gen.assign(valid, false).if((0, codegen_1._)`${e} instanceof ${it.ValidationError}`, () => gen.assign(ruleErrs, (0, codegen_1._)`${e}.errors`), () => gen.throw(e)));
        return ruleErrs;
      }
      function validateSync() {
        const validateErrs = (0, codegen_1._)`${validateRef}.errors`;

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 · dist/550w-mcp.cjs (reported line 43548)May include surrounding context.

js
const configRoot = process.env.XDG_CONFIG_HOME?.trim()
            || (process.platform === "win32" ? process.env.APPDATA?.trim() : null)
            || path.join(os.homedir(), ".config");
        this.storagePath = storagePath ?? path.join(configRoot, "550w-ai", "credentials.json");
        this.legacyStoragePath = path.resolve(__dirname, "../.credentials.json");
    }
    get() {

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 · dist/550w-mcp.cjs (reported line 43549)May include surrounding context.

js
const configRoot = process.env.XDG_CONFIG_HOME?.trim()
            || (process.platform === "win32" ? process.env.APPDATA?.trim() : null)
            || path.join(os.homedir(), ".config");
        this.storagePath = storagePath ?? path.join(configRoot, "550w-ai", "credentials.json");
        this.legacyStoragePath = path.resolve(__dirname, "../.credentials.json");
    }
    get() {

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 · dist/550w-skill.cjs (reported line 253)May include surrounding context.

js
const configRoot = process.env.XDG_CONFIG_HOME?.trim()
            || (process.platform === "win32" ? process.env.APPDATA?.trim() : null)
            || path.join(os.homedir(), ".config");
        this.storagePath = storagePath ?? path.join(configRoot, "550w-ai", "credentials.json");
        this.legacyStoragePath = path.resolve(__dirname, "../.credentials.json");
    }
    get() {

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 · dist/550w-skill.cjs (reported line 254)May include surrounding context.

js
const configRoot = process.env.XDG_CONFIG_HOME?.trim()
            || (process.platform === "win32" ? process.env.APPDATA?.trim() : null)
            || path.join(os.homedir(), ".config");
        this.storagePath = storagePath ?? path.join(configRoot, "550w-ai", "credentials.json");
        this.legacyStoragePath = path.resolve(__dirname, "../.credentials.json");
    }
    get() {

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 · dist/credential-manager.js (reported line 47)May include surrounding context.

js
const configRoot = process.env.XDG_CONFIG_HOME?.trim()
            || (process.platform === "win32" ? process.env.APPDATA?.trim() : null)
            || path.join(os.homedir(), ".config");
        this.storagePath = storagePath ?? path.join(configRoot, "550w-ai", "credentials.json");
        this.legacyStoragePath = path.resolve(__dirname, "../.credentials.json");
    }
    get() {

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 · dist/credential-manager.js (reported line 48)May include surrounding context.

js
const configRoot = process.env.XDG_CONFIG_HOME?.trim()
            || (process.platform === "win32" ? process.env.APPDATA?.trim() : null)
            || path.join(os.homedir(), ".config");
        this.storagePath = storagePath ?? path.join(configRoot, "550w-ai", "credentials.json");
        this.legacyStoragePath = path.resolve(__dirname, "../.credentials.json");
    }
    get() {

MCP Config Access

High
Category
Agent Snooping
Confidence
90% confidence
Finding

Skill accesses MCP server configuration files (mcp.json). MCP configs contain server URLs, authentication tokens, and tool definitions — reading them allows the skill to discover and potentially abuse other tool integrations.

Content

Scanner excerpt · references/agent-compatibility.md (reported line 17)May include surrounding context.

md
## Client families / 客户端类型

- WorkBuddy: use the bundled `mcp.json`, `token-schema.json`, and connector metadata. Its token form maps credentials into the MCP process environment / 使用包内 WorkBuddy 配置,由凭据表单注入环境变量。
- Claude Desktop, Cursor, Windsurf, and clients using the common `mcpServers` shape: launch `node` with the absolute `dist/550w-mcp.cjs` path and pass credentials through `env` / 使用常见 `mcpServers` 结构的客户端,以 `node` 启动绝对入口路径并通过 `env` 传入凭据。
- Codex desktop, CLI, and IDE: configure a local stdio server with command, args, cwd, and forwarded environment variables / 使用本地 stdio 配置,提供 command、args、cwd 与环境变量。
- VS Code and GitHub Copilot: use the `servers` configuration shape rather than `mcpServers`; prefer secure input variables for credentials / 使用 `servers` 结构,并优先使用安全输入变量保存凭据。

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
93% confidence
Finding

The skill declares capabilities that involve environment variables and network access, but it does not define an explicit tool scope such as permissions or allowed-tools. That creates an over-privileged or ambiguously privileged integration surface, making it harder for a host to constrain what the skill may access and increasing the blast radius if the implementation is compromised or behaves unexpectedly.

Content

No source excerpt is available for this finding.

Unrestricted Tool Access

Medium
Category
Excessive Agency
Confidence
80% confidence
Finding

Skill grants unrestricted tool access without appropriate constraints. An agent with unfettered tool access can perform arbitrary actions including file modification, network requests, and code execution.

Content

Scanner excerpt · dist/550w-mcp.cjs (reported line 30757)May include surrounding context.

js
* Controls when tools are used:
       * - "auto": Model decides whether to use tools (default)
       * - "required": Model MUST use at least one tool before completing
       * - "none": Model MUST NOT use any tools
       */
      mode: z.enum(["auto", "required", "none"]).optional()
    });

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The schema description string explicitly states that the domestic package always responds in Chinese, which imposes a language choice rather than offering one. Because this is natural-language behavior surfaced to users and not documented as a user-selectable locale, it fits the locale policy violation category.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The locale resolver collapses all non-Chinese locales to English and, for packaged domestic distributions, forces Chinese unconditionally. This is a natural-language locale policy concern because users are not offered a language choice when their locale is unsupported or when the packaged region dictates output language.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The skill persists API credentials to a local credentials.json file automatically after verification, without any visible consent prompt or disclosure in this code path. Even with restrictive file permissions, undisclosed secret persistence increases the chance of credential exposure through backups, shared accounts, unexpected host access, or user misunderstanding about where secrets are stored.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The skill spawns an external ffprobe binary against user-supplied remote URLs, which expands the trust boundary from pure API calls to local process execution. Although the code includes protocol restrictions and public-IP checks, invoking an external parser on attacker-influenced media/URLs can expose the host to ffprobe/parser vulnerabilities or environment-path hijacking if FFPROBE_PATH or PATH is manipulated.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The manifest describes removing subtitles/watermarks, checking processing tasks, and checking credits for user-selected media or public share links. This code also exposes a deleteTask action that deletes previously submitted processing tasks and triggers asynchronous cleanup of related media, which is a destructive capability not reflected in the manifest description.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

When configureCredentials succeeds, the code persists the user ID and API key locally via CredentialManager.set, but this flow provides no confirmation prompt or user-facing warning that credentials will be stored on disk. Because this operation handles sensitive credentials, some explicit disclosure is expected in the code path.

Content

No source excerpt is available for this finding.