The Agent Skills Directory

[OBFUSCATION]: The skill instructions explicitly teach the agent how to circumvent security tools. It directs the agent to use the ${VAR:+x} Bash expansion to bypass 'secret-expansion analyzers' and mandates splitting jq and curl commands into separate steps to evade 'Bash permission analyzers' that monitor network activity.
[COMMAND_EXECUTION]: The skill relies on the execution of a local script or binary named ./secretcurl to perform network POST requests. The source code and safety of this binary are not provided within the skill, posing a risk of executing unverified code on the host system.
[REMOTE_CODE_EXECUTION]: The combination of a custom network wrapper (./secretcurl) and the ingestion of remote payloads from an external API (https://api.x.ai/v1/responses) constitutes a risk where the local binary could be used to execute or download further malicious content if compromised or designed with backdoors.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from X (Twitter) and general web searches to perform 'position calls' and narrative mapping.
Ingestion points: External data enters the context via the X.AI API responses (stored in /tmp/xai-nt.json) and WebSearch results.
Boundary markers: None are present; the skill does not use delimiters or instructions to ignore potential commands embedded in the fetched narratives.
Capability inventory: The skill has the ability to perform network POSTs via ./secretcurl, write to the repository (memory/logs/ and memory/MEMORY.md), and send notifications via ./notify.
Sanitization: The skill uses jq for structural parsing but performs no sanitization or filtering of the natural language content before it is used to influence the agent's logic and positioning decisions.

narrative-tracker