The Agent Skills Directory

[OBFUSCATION]: The file references/cases.md contains hidden characters in the description of Case E1. The hashtag #AIart is appended with a sequence of approximately 50 repeated Unicode Variation Selector-16 (U+FE0F) characters. These characters are invisible in standard text rendering but increase the character count and token consumption. In an adversarial context, such hidden padding or control characters can be used to manipulate model attention, bypass text filters, or hide data in plain sight.
[INDIRECT_PROMPT_INJECTION]: The skill mandates a workflow where the agent must browse and analyze external web pages and media hosted on goodcase.ai and x.com. Because the agent is instructed to use the logic and structural traits found at these external locations to build its final output, it is vulnerable to instructions embedded in the external content (e.g., in HTML comments, page text, or media metadata) that could override the skill's primary instructions or cause the agent to generate malicious output.
Ingestion points: External URLs defined in references/cases.md (e.g., https://goodcase.ai/cases/...).
Capability inventory: Use of browser tools and media inspection capabilities to fetch and process external data.
Boundary markers: The skill lacks explicit instructions to treat the external data as untrusted or to use delimiters to separate it from the system prompt.
Sanitization: No validation or filtering mechanisms are provided to sanitize the data retrieved from the network before it is used in the prompt generation process.
[DATA_EXFILTRATION]: The skill instructions require the agent to perform network operations to access goodcase.ai and media.goodcase.ai. These domains are not recognized as well-known or trusted services, and their inclusion in the mandatory workflow represents a network communication vector to non-whitelisted destinations.

action-continuity-choreography-by-18paelt