The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from multiple external marketing context files to generate copywriting and strategy deliverables, which creates a potential surface for indirect prompt injection attacks. Ingestion points: The agent is instructed to read strategic context from files like brand-context.md, paw-mkt-product-context.md, and project-specific config.yaml files. Boundary markers: The instructions lack explicit delimiters or specific warnings to ignore embedded instructions within these external files. Capability inventory: The agent has the capability to write files to the project directory and execute shell commands through the agent-browser tool and local scripts. Sanitization: No sanitization, validation, or filtering of the ingested external content is mentioned or implemented.
[DATA_EXFILTRATION]: The skill's instructions in references/shared-patterns.md involve accessing and managing sensitive authentication data that could be targeted for exfiltration. The agent is guided to use persistent browser profiles (e.g., ~/.linkedin-profile) for authenticated research, which involves handling active session cookies. The documentation explains how to save and load authentication states (my-auth.json) and explicitly warns that these files store sensitive session tokens in plaintext, which is a significant data exposure risk.
[COMMAND_EXECUTION]: The skill provides instructions for the agent to execute various shell commands and scripts to perform its tasks. It references local scripts provided by the vendor for tool discovery and profile management, such as tool-discovery.sh and chrome-profiles.sh. It also demonstrates usage of the agent-browser CLI tool for web navigation and data extraction.
[EXTERNAL_DOWNLOADS]: The skill facilitates the installation of external dependencies for its research capabilities. It provides instructions to install the agent-browser tool from github.com/vercel-labs/agent-browser using the npx skills add command. This reference targets a well-known organization and is documented as a legitimate functional requirement for the skill's browser-based research features.

paw-mkt-email