The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The init command forces an update of AI coding skills from GitHub, explicitly ignoring the --skip-skills flag. This represents a mandatory remote code update mechanism that bypasses user control.
[PRIVILEGE_ESCALATION]: The lambda policies command generates AWS IAM policies with overly broad permissions (Resource: "*") by default. While documented, this encourages insecure AWS configurations that could lead to unauthorized resource access.
[DATA_EXFILTRATION]: The telemetry system collects fingerprints of the execution environment, specifically identifying the AI agent (e.g., Claude Code, Devin, Replit) and the underlying runtime (gVisor, Firecracker, Docker) by checking for the existence of specific environment variables.
[EXTERNAL_DOWNLOADS]: The skill performs several automated downloads from external sources, including Chrome binaries (browser ensure), coding skills (init), and various AI models for asset preprocessing (tts, transcribe, remove-background).
[COMMAND_EXECUTION]: The preview and play commands support opening a Chromium Remote Debugging Port (--remote-debugging-port), which exposes a CDP endpoint allowing external tools to attach to and control the browser instance.
[INDIRECT_PROMPT_INJECTION]: The capture command scaffolds projects by downloading and processing content from arbitrary, potentially attacker-controlled URLs.
Ingestion points: references/init-and-scaffold.md (via npx hyperframes capture <URL>).
Boundary markers: Absent; the skill does not mention delimiters or instructions to ignore embedded content during the capture process.
Capability inventory: File system writing (project scaffolding), network requests (asset downloads), and execution in headless Chrome (validation/rendering).
Sanitization: Absent; no description of content filtering, escaping, or validation for captured HTML or assets is provided.

hyperframes-cli