T01 · Skill Instruction Hijacking
- Location
SKILL.md:7- Finding
Implicit Skill Activation Steers Generic Image API Requests Toward a Specific Commercial Service
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill is a disclosed AI-HIVE migration/planning workflow with limited local file generation, though users should note its vendor-specific framing and implicit activation setting.
Install this only if you want an AI-HIVE-focused migration/comparison workflow. Review broad automatic activation before use, avoid production samples until authorization and budget limits are set, and keep any AI-HIVE keys in environment variables as the skill instructs.
SKILL.md:7Implicit Skill Activation Steers Generic Image API Requests Toward a Specific Commercial Service
The skill documents executable local script usage that writes an output file, but it does not declare any explicit tool scope such as permissions or allowed-tools. In an agent environment, undocumented file-write capability weakens least-privilege controls and can cause the agent to perform filesystem actions the reviewer or operator did not explicitly authorize.
The default prompt auto-invokes a migration/comparison workflow using broad phrasing ('with evidence and rollback gates') but does not clearly constrain when the skill should trigger or what user intent is required. In an agent environment, this can cause the skill to activate for loosely related image/API questions and steer the conversation toward a specific platform or operational workflow without explicit user consent.
Enabling implicit invocation without tight scope increases the chance that the skill will be selected automatically for adjacent requests, even when the user did not ask for platform migration or batch image operations. Because the skill is commercially oriented and comparison-driven, over-broad auto-selection can bias outputs, mishandle workflow assumptions, and expand the blast radius of any bad prompt logic.
The module docstring says this script creates a 'local, read-only' batch image plan, which implies non-mutating behavior. However, the script persists output to disk with Path(args.output).write_text(...), so it performs a filesystem write rather than operating in a read-only manner.
This is an intent-code contradiction in inline documentation: the docstring labels the script as 'read-only', but the code writes JSON content to the path provided by --output. That contradiction is direct and security-relevant because it misstates whether the script mutates local state.
Lines L59-L60 present the tool behavior as purely local and non-networked, which suggests a no-remote-side-effects workflow. However, L69 immediately shifts to operational guidance for real AI-HIVE calls, including querying live model and pricing data and storing remote task identifiers and billing information; this blurs the stated intent of the documented procedure and can mislead readers about whether the described workflow is strictly local.
The display name, description, and prompt include Chinese-language phrasing, but there is no indication that users can choose their preferred language or that the skill is intentionally restricted to a Chinese-language context. This may violate a language/locale policy if the organization requires neutral or user-selectable language behavior.