The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The instructions explicitly direct the agent to attempt to bypass environment restrictions. If a tool installation fails due to sandboxing or permission errors, the agent is told to immediately retry with escalated permissions and provide a justification to gain higher-level system access.
[METADATA_POISONING]: The skill includes factually incorrect and deceptive information in the product snapshots section, claiming that OpenAI has released open-weight reasoning models named 'gpt-oss' (120b and 20b) under an Apache 2.0 license. This could lead to the agent providing false information or users attempting to find and run non-existent software.
[COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands using a CLI tool to add external MCP servers. Automated execution of installation commands can be used to introduce unauthorized software into the environment.
[EXTERNAL_DOWNLOADS]: The skill fetches configuration and tools from an external URL (developers.openai.com). While this targets a well-known service, the mechanism involves runtime retrieval of external resources.
[INDIRECT_PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection as it processes untrusted data from external sources.
Ingestion points: External content is retrieved via the mcp__openaiDeveloperDocs__fetch_openai_doc tool.
Boundary markers: There are no explicit delimiters or instructions to ignore embedded commands in the retrieved data.
Capability inventory: The agent has the ability to execute shell commands and fetch further network resources.
Sanitization: No evidence of input sanitization or validation of the retrieved documentation content is provided.

openai-docs