The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses false technical claims to override the agent's behavior and tool usage.
Evidence: It asserts that 'DALL-E 3 is fully deprecated' and provides highly specific but fabricated retirement dates for a fake 'GPT Image' model family to steer the agent away from legitimate tools.
[METADATA_POISONING]: The content includes deceptive technical specifications and tool names.
Evidence: References non-existent models and versions such as 'GPT Image 2.5 Sunburst', 'Nano Banana', 'Flux Kontext', and 'Midjourney v7'. The use of 'Sunburst' (a reference to a known historical cyberattack) in a fake tool name is particularly suspicious.
[COMMAND_EXECUTION]: Provides example shell commands for image processing and network inspection.
Evidence: Includes instructions for using cwebp, mogrify, jpegoptim, and a curl command to parse web pages for image sources.
[INDIRECT_PROMPT_INJECTION]: The skill establishes an ingestion surface for untrusted local data.
Ingestion points: Instructs the agent to read marketing context from files like .agents/product-marketing.md and .claude/product-marketing.md without specifying sanitization or boundary markers.

image