The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection as it processes untrusted external data.
Ingestion points: In Step 1, the skill requests a "product page link" and "2–5 photos" which are processed by the agent and the arcads_analyze_media tool.
Boundary markers: Absent. The instructions do not specify the use of delimiters or clear "ignore instructions" markers when handling content extracted from external URLs.
Capability inventory: The skill possesses extensive media generation capabilities via arcads_generate_video_seedance_25, arcads_generate_image_gpt, and arcads_text_to_speech. An attacker could theoretically influence the output of these tools by embedding malicious instructions in the product page.
Sanitization: The skill provides logic to extract specific data (product name, benefits, ingredients), which serves as a basic filter, but it does not specify rigorous sanitization or escaping of the extracted text before it is used in subsequent prompts.
[SAFE]: The skill uses a suite of custom tools (e.g., arcads_analyze_media, arcads_generate_video_seedance_25, arcads_stitch_videos) that are consistent with the developer's identity (arcads-ai) and the intended purpose of the skill. No unauthorized tool access or privilege escalation attempts were found.
[SAFE]: Network operations are restricted to the primary function of fetching product information from user-supplied links. The skill includes explicit instructions for the agent to fail gracefully if the link is blocked (403) and prohibits alternative fetching methods, which reduces the risk of SSRF or unintended network probing.

cinematic-ads