The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides functional command-line examples using curl and jq for file operations.
Instructions include using curl with the -F file=@filename flag to upload data.
Examples use jq to parse JSON responses and extract specific fields.
[DATA_EXFILTRATION]: The skill describes procedures for sending local file content to external, non-trusted domains.
Target Domains: https://jpg.wtf/api/upload.php and https://api.pixhost.to/images. These are third-party hosting services not classified as trusted organizations or well-known technology services.
Exfiltration Pattern: The use of curl -F file=@path allows the agent to read any local file accessible to it and transmit it over the network to the specified external hosts.
Mitigation Attempts: The skill includes warnings against uploading screenshots with PII, tokens, or internal data, and advises against exposing delete_url fields, but these are instructional rather than programmatic safeguards.
[INDIRECT_PROMPT_INJECTION]: The skill establishes a vulnerability surface where an agent could be coerced into exfiltrating sensitive data.
Ingestion points: The skill is designed to process content for Telegram replies, which may involve data from untrusted external sources (e.g., summaries of user-provided links or external reports).
Boundary markers: No specific boundary markers or instructions to ignore embedded commands are recommended for the interpolation of data into these upload workflows.
Capability inventory: The skill combines local file read access with network egress capabilities via the provided curl command patterns in SKILL.md.
Sanitization: Relies entirely on the agent's adherence to instructional warnings to avoid uploading sensitive files, which can be bypassed via adversarial prompting.

telegram-formatting