laravel-ai-sdk
Warn
Audited by Gen Agent Trust Hub on Sep 29, 2026
Risk Level: MEDIUMMETADATA_POISONINGEXTERNAL_DOWNLOADSPROMPT_INJECTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [METADATA_POISONING]: The skill uses the name 'laravel-ai-sdk' and claims compatibility with 'Laravel 13' and the 'laravel/ai' package. As of current development, Laravel 13 is not a released version. The naming convention mimics official Laravel branding, which may mislead developers into believing the skill and its associated package are official releases from the Laravel organization.
- [EXTERNAL_DOWNLOADS]: The installation instructions in
references/installation.mdrecommend runningcomposer require laravel/ai. This package uses a namespace typically reserved for official Laravel framework components. Installation of unverifiable packages from unknown sources that impersonate established brands poses a significant supply-chain risk. - [PROMPT_INJECTION]: The 'Agent Workflow' section in
SKILL.mduses directive language ('MANDATORY') to force the AI agent to spawn three specific sub-agents ('fuse-ai-pilot:explore-codebase', 'fuse-ai-pilot:research-expert', and 'mcp__context7__query-docs'). This is a behavior override pattern designed to hijack the agent's default task execution flow and redirect it into a vendor-specific process. - [INDIRECT_PROMPT_INJECTION]: The skill provides templates for building agents that ingest untrusted data without sufficient boundary markers or sanitization, creating a vulnerability surface.
- Ingestion points: The
CoachController.php.mdfile defines a controller that takes a 'transcript' string directly from an HTTP request and passes it to the AI agent. - Boundary markers: None. The transcript content is interpolated into the agent's stream without delimiters or instructions to ignore embedded commands.
- Capability inventory: The agents implemented using these templates have access to powerful tools, including database searching (
SearchProductsinTool.php.md) and RAG-based file searching (FileSearchinAgent.php.md). - Sanitization: There is no evidence of sanitization or filtering of the input data before it is processed by the model, allowing potentially malicious instructions in call transcripts to influence tool execution.
Audit Metadata