The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest data from diverse and untrusted external sources including Slack, Jira, Confluence, emails, and the public web. There is a risk that malicious instructions embedded in these sources could influence agent behavior during processing.
Ingestion points: Jira issues, Confluence wikis, Slack chat, Mail, Calendar, source repositories, and external web content.
Boundary markers: The skill specifies strict formatting and headers for the report, but does not include explicit prompt delimiters or 'ignore' instructions to prevent the agent from executing content found within verbatim quotes.
Capability inventory: The skill utilizes subagents and has broad read access across internal and external systems, as well as the ability to write reports to the local file system.
Sanitization: No sanitization is mentioned; instructions explicitly require quoting sources 'verbatim', which can lead to the interpretation of data as instructions.
[DATA_EXFILTRATION]: The skill targets an exceptionally broad range of sensitive data environments including 'Cost and billing', 'Mail, calendar, and meeting records', 'Cloud runtime state', and 'Source repos'. While the instructions mandate read-only operations and prohibit printing secrets, the aggregation of this data into a single timeline represents a significant exposure risk.
[COMMAND_EXECUTION]: The skill uses dynamic discovery to 'Follow the evidence', fanning out subagents to newly discovered sources (e.g., a ticket naming a repo or a message naming a mailbox). This recursive tool invocation dynamically expands the agent's reach based on potentially untrusted input content.

archeology