The Agent Skills Directory

[METADATA_POISONING]: The skill contains extensive documentation for non-existent Anthropic models, including Claude Fable 5, Claude Mythos 5, Claude Opus 5, and Claude Sonnet 5. It includes specific, hallucinated technical details, pricing, and release dates (June 2026). This misleading content can cause AI agents to misinterpret their environment and attempt to use non-existent API features or model IDs.
[EXTERNAL_DOWNLOADS]: The skill provides commands to download the ant CLI tool directly from github.com/anthropics and install various SDKs from official package registries like npm, PyPI, and Maven. These external references target official and trusted vendor repositories.
[COMMAND_EXECUTION]: The documentation includes numerous shell command examples for the ant CLI and curl to interact with the Claude API. These include authentication, message creation, and model management tasks.
[PRIVILEGE_ESCALATION]: The installation script for the ant CLI tool utilizes sudo tar to install the binary into /usr/local/bin, a protected system directory.
[INDIRECT_PROMPT_INJECTION]: The skill introduces a high risk of indirect injection by providing an agent with a large set of false technical specifications and documentation. An agent trusting this skill will operate under a false context regarding the current state of Claude models.
Ingestion points: Documentation files in references/ and scripts/ folders.
Boundary markers: None identified; instructions are presented as authoritative facts.
Capability inventory: The skill encourages use of shell execution via CLI tools and package managers.
Sanitization: None present.

anthropic-models-sdks