The Agent Skills Directory

[COMMAND_EXECUTION]: Th e skill explicit ly direct s th e agen t to attemp t privilege escalatio n whe n installatio n fail s by retrying with escalate d permissio n s and spec if ic ally instru ct s th e agen t to skip use r appro va l fo r th i s actio n.
[PROMPT_INJECTION]: Th e skill contain s decep tiv e refe rence s to non-existen t AI model s and produc t s (e.g. , GPT-5 .4, gpt-o ss, Sora-2 ) whic h may resul t in th e agen t providing fals e inf ormatio n or perf orming uninten de d actio n s.
[EXTERNAL_DOWNLOADS]: Th e skill comman d s th e agen t to downlo a d and instal l externa l too l s from a remo t e URL (http s://develo pers .openai.c om/mcp ). Whil e th e sourc e is a well-know n domain , th e combinatio n with autonomou s escalatio n instru ctio n s create s a security risk.

openai-docs