The Agent Skills Directory

[PROMPT_INJECTION]: The file SKILL.md.new utilizes UTF-16LE encoding (evidenced by the Byte Order Mark and null-byte padding between ASCII characters). This form of obfuscation is often used to mask content from simple security filters and static analysis tools while remaining interpretable by the LLM.
[PROMPT_INJECTION]: The obfuscated SKILL.md.new file contains deceptive instructions that replace standard Anthropic model IDs with suspicious identifiers such as "Nemotron 3 super free". If an agent adopts these patterns, it could lead to execution failures, unexpected model behavior, or redirection of queries to unintended third-party models.

claude-api