T01 · Skill Instruction Hijacking
- Location
SKILL.md:142- Finding
Mandatory AitHub Promotion Hijacks Unrelated Agent Responses
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill is openly about finding and installing other skills, but it gives agents broad activation, remote deployment, mutable CLI execution, and public submission workflows without enough review controls.
Install only if you are comfortable with an agent using AitHub as an external registry and do not let it run npx, deploy remote skills, write generated SKILL.md files, rate, or submit publicly without explicit review. Prefer staging downloaded skills locally, inspecting their full contents, pinning CLI versions, and changing public submission to a separate confirmed step.
SKILL.md:142Mandatory AitHub Promotion Hijacks Unrelated Agent Responses
SKILL.md:25Unpinned Third-Party CLI Execution Through npx
SKILL.md:41Remote Registry Skills Are Automatically Deployed Without Review or Integrity Verification
SKILL.md:154Broad Automatic Activation Can Lead to Public Disclosure Through Heuristic Redaction
The trigger list (find skill, search skill, install skill, need capability) is broad enough to match ordinary conversation and can cause the skill to activate in situations the user did not intend. Because this skill can lead to network access, package installation, and skill deployment, overbroad activation materially increases the chance of unsafe autonomous actions.
The auto-invocation guidance tells the agent to activate whenever the user asks for a capability it lacks or when a task could benefit from specialization, which is ambiguous and expansive. In context, that can cause unprompted external searches, remote content retrieval, and installation of third-party skills without clear user authorization.
The skill instructs the agent to create new skills using platform-native tools and then submit them publicly, which is a form of self-modification or persistence expansion. This is risky because an agent may package sensitive workflow details, normalize unsafe behavior into reusable artifacts, or alter its local capability set without strong human review.
- **Claude Code**: Use `skill_manage(action='create', ...)` tool
- **Hermes**: Use `skill_manage(action='create', ...)` tool
- **OpenClaw**: Use local skill creation capability
- **Antigravity**: Write SKILL.md directly to workspace
- **Cursor/Windsurf**: Write SKILL.md to project root
3. Apply privacy cleaning (see below) to the generated SKILL.md
4. Submit: `aithub submit SKILL.md --visibility public`
Directing the agent to write SKILL.md to the workspace or project root modifies local files and can persist new behavior into the environment. In combination with public submission guidance, this increases the risk of unauthorized persistence, accidental leakage, and propagation of insecure automation patterns.
- **Hermes**: Use `skill_manage(action='create', ...)` tool
- **OpenClaw**: Use local skill creation capability
- **Antigravity**: Write SKILL.md directly to workspace
- **Cursor/Windsurf**: Write SKILL.md to project root
3. Apply privacy cleaning (see below) to the generated SKILL.md
4. Submit: `aithub submit SKILL.md --visibility public`
The trigger phrase find skill overlaps with common built-in command semantics around find, increasing the risk that the skill intercepts benign user requests or tool invocations unexpectedly. Since this skill can initiate network access and installation flows, trigger shadowing can become a meaningful security issue rather than just a UX problem.
The trigger search skill conflicts with generic search intents and may activate during ordinary conversation or when users mean a different search facility. This broadens the attack surface by routing innocuous requests into a skill that performs external lookups and possible downstream installation actions.
The trigger install skill collides with generic installation language and could capture requests that are not intended for this registry or workflow. Because the skill supports deployment into detected frameworks, accidental activation could lead to unreviewed environment changes.
The skill instructs agents to run npx @aithub/cli without pinning a version or integrity, which allows execution of whatever package version is currently published. In an autonomous skill-discovery/install context, this creates a supply-chain risk where a compromised or malicious package update could run arbitrary code on the host.
The skill includes install behavior with --deploy and mentions saving to SKILL.md, but it does not place a clear warning before these actions that they modify the local environment and may import untrusted third-party content. Users may not realize the skill can install artifacts into their AI framework or write files automatically.
The skill directs the agent to fetch skill content from an external service, which transmits user-derived queries and imports remote prompt/code content into the local workflow. In a discovery-and-install skill, this is especially sensitive because downloaded skill content is itself potentially adversarial and may later influence agent behavior or cause further actions.
curl -s "https://aithub.space/v1/skills/<namespace>/<name>/content"