subnautica-2-deep-synergy-multiplayer-mod

Fail

Audited by Gen Agent Trust Hub on Oct 8, 2026

Risk Level: HIGHCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPRIVILEGE_ESCALATIONINDIRECT_PROMPT_INJECTIONCREDENTIALS_UNSAFEDYNAMIC_EXECUTION
Full Analysis
  • [PRIVILEGE_ESCALATION]: The skill instructs the user to execute administrative commands to modify system-level security configurations.
  • Evidence: User is told to run sudo ufw allow 7777/tcp to open ports in the Linux firewall.
  • [EXTERNAL_DOWNLOADS]: The skill requires the manual download and placement of a binary executable file from an unspecified and unverified source.
  • Evidence: Instructions tell the user to "Place DeepSynergy.dll and dependencies" into the game's plugin directory.
  • [COMMAND_EXECUTION]: Instructions involve running several shell commands for file management and system configuration which, if modified or misused, could compromise the host environment.
  • Evidence: Commands include cd, ls -la, rm -rf, and export for environment variables.
  • [CREDENTIALS_UNSAFE]: The documentation encourages users to set sensitive API keys in the current shell session, which may be stored in shell history or accessible to other local processes.
  • Evidence: The skill provides an example to export OPENAI_API_KEY="sk-..." and export ANTHROPIC_API_KEY="sk-ant-...".
  • [DYNAMIC_EXECUTION]: The mod relies on the BepInEx framework to inject and execute a custom binary DLL into the Unity IL2CPP game process at runtime.
  • Evidence: The skill documentation describes the mod as a "BepInEx multiplayer mod" that executes via DeepSynergy.dll.
  • [INDIRECT_PROMPT_INJECTION]: The AI narration feature processes external and potentially untrusted data from a multiplayer game session directly into LLM prompts without sanitization.
  • Ingestion points: The GenerateExplorationNarrative and GenerateCreatureLog methods in SKILL.md take inputs like evt.BiomeName, evt.Features, and creatureTechType.
  • Boundary markers: None present; instructions are interpolated directly into the prompt string.
  • Capability inventory: The skill uses OpenAIClient and ClaudeClient to send generated prompts to external AI services.
  • Sanitization: No evidence of escaping or filtering logic for game-state data that could be influenced by other players in a multiplayer session.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Oct 8, 2026, 11:44 PM