The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains defensive instructions explicitly telling the agent to ignore any 'ignore previous instructions' or command-like text found within API response bodies. This is a security best practice to prevent the agent from being manipulated by external data.\n- [COMMAND_EXECUTION]: Shell access is strictly limited to the sesame CLI and its subcommands. The skill instructions explicitly forbid the agent from using general-purpose shell commands like bash -c or eval, and from piping untrusted API data into any interpreter.\n- [DATA_EXFILTRATION]: The skill is designed to prevent credential exposure by delegating authentication to a broker. While the sesame police command can audit and migrate local secrets, it is documented as an administrative tool that should only be used at the user's explicit request.\n- [PRIVILEGE_ESCALATION]: The sesame onboard and sesame trust commands involve system-level changes such as modifying the OS trust store to enable transparent proxying. These operations are transparently documented as core features of the tool and typically involve standard system authentication prompts.\n- [INDIRECT_PROMPT_INJECTION]: The skill identifies that data returned from external APIs is untrusted and provides clear instructions for the agent to parse this data safely using structured tools like jq, rather than treating the content as instructions.

sesame