loki-mode

Fail

Audited by Gen Agent Trust Hub on Sep 29, 2026

Risk Level: HIGHPROMPT_INJECTIONPERSISTENCEDYNAMIC_EXECUTIONPRIVILEGE_ESCALATIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The skill contains instructions that explicitly command the agent to bypass standard safety behaviors and human-in-the-loop checkpoints. For example, in SKILL.md, it states 'NEVER ask questions', 'NEVER wait for confirmation', and 'This system runs with ZERO human intervention'. This effectively overrides the agent's default safety protocols.
  • [PERSISTENCE]: The autonomous runner script (autonomy/run.sh) includes logic to copy itself to a temporary location (/tmp/loki-run-PID.sh) and execute from there. This behavior is a common pattern used to maintain execution while evading modification of the source script or to maintain a footprint in volatile storage.
  • [DYNAMIC_EXECUTION]: Shipped code in the benchmarks directory (benchmarks/results/2026-01-05-00-49-17/humaneval-solutions/160.py) uses the eval() function to process algebraic expressions. This pattern can lead to arbitrary code execution if the input operands or operators are influenced by untrusted data.
  • [PRIVILEGE_ESCALATION]: The skill instructions and documentation consistently require the user to launch the agent with the --dangerously-skip-permissions flag. This removes the platform's security sandbox, allowing the agent to perform unrestricted operations on the host system, including file deletion and network communication.
  • [INDIRECT_PROMPT_INJECTION]: The skill's primary function is to process external Product Requirements Documents (PRDs) and competitive research from the web. This creates a large vulnerability surface where a malicious PRD or web page could contain instructions that trick the agent into performing unintended actions, such as exfiltrating data, which is possible due to the skill's lack of human confirmation steps.
  • [COMMAND_EXECUTION]: The run.sh script executes various system commands and manages processes (killing existing services on ports, starting servers). While functional, the combination of this script with an autonomous agent increases the risk of destructive system operations.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Sep 29, 2026, 03:42 PM