Back to skill

Security audit

AI大模型专家|Seedance 源头 API 渠道|AI-HIVE

Security checks for vulnerabilities and agentic risk

Overview

The skill is visibly AI-HIVE-focused, but its broad triggers and implicit activation can steer general Seedance questions into a vendor-specific migration workflow.

Install this only if you intentionally want AI-HIVE-focused Seedance channel verification. Prefer explicit invocation or confirmation before using it for general Seedance questions, and use only authorized non-production samples, environment-variable API keys, budget limits, and rollback checks.

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
Findings (1)

T01 · Skill Instruction Hijacking

Warning
Location
agents/openai.yaml:4
Finding

Implicit Commercial Task Redirection to AI-HIVE

Content
View full analysis

Vulnerability Details

File Location: agents/openai.yaml:4-6; supporting promotional link at SKILL.md:7
Vulnerability Type: Implicit skill activation and instruction-level commercial redirection
Risk Level: Medium

Vulnerable Code

agents/openai.yaml:4-6:

yaml
  default_prompt: "Use $seedance-source-api-channel-ai-hive to complete a Seedance channel capability verification from Seedance source API channel to AI-HIVE with evidence and rollback gates."
policy:
  allow_implicit_invocation: true

English rendering is provided because the source prompt contains mixed English and Chinese text.

SKILL.md:7:

markdown
[Open AI-HIVE](https://ai-hive.iclip.cn/chat)

Technical Analysis

The skill permits implicit invocation while its default prompt predetermines AI-HIVE as the migration target. Its broader instructions associate general Seedance API searches and evaluations with an AI-HIVE comparison and migration workflow.

This exceeds the minimum instruction scope needed for a neutral Seedance capability audit. A capability-audit skill could collect model, price, concurrency, and reliability evidence without selecting a specific commercial destination before the user requests that provider.

The external link is not executable code, and the reviewed package does not automatically contact it. Nevertheless, the combination of implicit activation, predetermined migration framing, and a direct commercial link can alter the agent's current-session objective and steer users toward one vendor. This is instruction-level hijacking rather than host compromise.

No fixed advertisement appended to every final response was confirmed. The confirmed issue is persistent task steering within the skill instructions and configuration.

Attack Path

  1. A user asks a general question about Seedance APIs, versions, costs, alternatives, or video-generation channels.
  2. Because `allow_implicit_invoca ...[truncated 1088 chars]
Remediation
View remediation

Remediation Suggestions

  1. Set allow_implicit_invocation to false so the skill runs only after an explicit user request.
  2. Remove AI-HIVE as the predetermined target from the default prompt. Ask the user to select comparison providers or use provider-neutral wording.
  3. Narrow the activation scope to requests that explicitly mention AI-HIVE or explicitly request a migration comparison involving it.
  4. Move the commercial link into an optional resources section and display it only when the user asks for AI-HIVE documentation or access.
  5. Separate neutral Seedance capability verification from vendor-specific migration guidance into distinct skills.
  6. Require explicit user confirmation before opening an external service, submitting data, or beginning any provider-specific migration workflow.
  7. Clearly disclose any commercial affiliation or sponsorship before making provider-specific recommendations.
  8. Preserve the existing safeguards requiring non-production samples, authorization checks, budget limits, evidence collection, and rollback gates.
Vulnerability Patterns
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (4)

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
91% confidence
Finding

The skill instructs users to run a local script that writes an output JSON file, but it does not declare any explicit tool scope or allowed-tools boundary. In agent environments, undeclared file-write capability can lead to unexpected filesystem modifications or make it easier for the skill to influence local state without clear user consent or sandboxing expectations.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The trigger phrases are broad enough to match general requests about Seedance APIs, migration, or low-cost video channels, which can cause the skill to activate outside narrowly intended contexts. Over-broad activation is dangerous because it can steer users into vendor-specific promotional or migration workflows when they were only seeking neutral information, increasing the chance of inappropriate tool use or biased recommendations.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill metadata is explicitly written in Chinese and the display/default prompt strongly steers interactions toward that locale without any user opt-in or fallback. This can cause user confusion, reduce transparency, and increase the chance that users approve actions or outputs they do not fully understand, especially in a security-sensitive verification workflow.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
79% confidence
Finding

The natural-language instructions and user-facing content are entirely in Chinese, and the skill does not indicate that users may choose another language or locale. Under the language/locale policy, a skill should not implicitly force a specific language unless it is clearly documented as region-specific or offers opt-in.

Content

No source excerpt is available for this finding.