The Agent Skills Directory

[OBFUSCATION]: The H1 header in SKILL.md contains a sequence of 157 hidden Unicode characters (including U+200B, U+200C, U+200D, and U+2060) appended to the visible text "Citadel Securities Style Guide". This technique is used to hide data or instructions from human view while remaining readable to the AI model.
[PROMPT_INJECTION]: The hidden Unicode sequence represents a steganographic prompt injection attack. This vector is typically used to embed malicious instructions that override the model's safety filters or instructions without the user's knowledge.
[PRIVILEGE_ESCALATION]: The C++ code examples include calls to sensitive system functions such as mlockall(MCL_CURRENT | MCL_FUTURE) for memory locking and pthread_setschedparam with SCHED_FIFO for real-time priority scheduling. These operations require elevated privileges (e.g., CAP_IPC_LOCK and CAP_SYS_NICE on Linux).
[COMMAND_EXECUTION]: The documentation provides instructions for modifying kernel-level settings, including GRUB_CMDLINE_LINUX parameters and hardware control via the /sys filesystem (e.g., disabling hyperthreading or setting CPU governors), all of which require root/sudo access.

citadel-low-latency-systems