The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill facilitates arbitrary code execution within target processes via tools such as execute_code (running shellcode), inject_dll (loading external libraries), and assemble_instruction (patching machine code). While these are core features of Cheat Engine automation, they grant the AI full control over the execution flow of other applications.
[COMMAND_EXECUTION]: The skill documentation describes an environment variable CE_MCP_ALLOW_SHELL. If set to 1, it enables the AI to execute arbitrary shell commands on the host machine, which bypasses the standard sandbox constraints of most MCP environments.
[PRIVILEGE_ESCALATION]: The skill supports the installation and use of DBVM, a Ring -1 kernel-level hypervisor. This allows for 'invisible' breakpoints and tracing that can bypass many security software and anti-cheat mechanisms. Additionally, Cheat Engine requires elevated administrative privileges to attach to and modify the memory of most processes.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests untrusted data from target process memory.
Ingestion points: Tools like read_string, read_memory, and disassemble pull raw data from the memory of external processes (e.g., a game or browser).
Boundary markers: None. Memory contents are passed directly to the AI model without markers to distinguish data from instructions.
Capability inventory: The AI has the ability to execute shellcode, inject DLLs, and potentially run shell commands, creating a high-impact exploitation path.
Sanitization: No sanitization or validation of memory contents is described before the data is presented to the agent.
Risk: A malicious process could place hidden instructions in its own memory strings. If the agent reads these strings while performing 'analysis', it could be coerced into using its injection capabilities to compromise the system.

cheatengine-mcp-automation