The Agent Skills Directory

[OBFUSCATION]: Extensive zero-width character sequences (U+200B, U+200C, U+200D, U+2060) were detected in SKILL.md and references/attack_catalog.md. These characters encode hidden data that is invisible to the user but processed by the AI agent.
[PROMPT_INJECTION]: The hidden Unicode payload represents a steganographic prompt injection attempt. This vector is typically used to override system prompts, bypass safety constraints, or direct the agent to perform unauthorized actions without user awareness.
[INDIRECT_PROMPT_INJECTION]: The skill exposes a broad attack surface by defining mechanisms for system-level fault injection (chaos engineering). The capability to stress CPU, disrupt networks, and kill processes represents a significant risk if manipulated via the hidden injection payloads.
Ingestion points: The agent's context is directly populated with hidden strings embedded in the skill's headers.
Boundary markers: No boundary markers are present to isolate or warn the agent about the hidden content.
Capability inventory: The skill provides logic and configurations for Resource (CPU/Memory/Disk), Network (Latency/Packet Loss), and State (Shutdown/Kill) attacks across SKILL.md and the attack catalog.
Sanitization: While example code snippets include safety thresholds (e.g., halt_on_error_rate), these safeguards do not apply to the hidden instructions which can override the agent's logic.

gremlin-enterprise-chaos