meridian-privilege-escalation

Fail

Audited by Gen Agent Trust Hub on Oct 23, 2026

Risk Level: HIGHPRIVILEGE_ESCALATIONPROMPT_INJECTIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [PRIVILEGE_ESCALATION]: The skill explicitly documents how to acquire maximum permissions and bypass sandbox restrictions.
  • Instructions provide syntax for --sandbox danger-full-access, which grants full filesystem and process access.
  • Instructions provide syntax for --approval yolo, which enables the agent to approve all tool calls without user confirmation or prompts.
  • The documentation describes switching models specifically to "escape a local harness restriction," which is a technique for bypassing security controls.
  • [PROMPT_INJECTION]: The skill contains instructions that encourage bypassing safety protocols.
  • It details how to "disable safety checks entirely" using broad overrides.
  • It provides a manual for an agent to escalate its own permissions when it encounters "constrained permissions" or "restricted tools."
  • [COMMAND_EXECUTION]: The skill demonstrates the execution of shell commands (meridian spawn) with parameters that grant unrestricted capability over the host environment.
  • [INDIRECT_PROMPT_INJECTION]: The skill establishes an attack surface for indirect injection by spawning subagents that process external files under elevated privileges.
  • Ingestion points: External task files referenced via --prompt-file task.md (SKILL.md).
  • Boundary markers: None identified; instructions do not suggest delimiters or warnings for the content of subagent prompt files.
  • Capability inventory: The spawned subagents can gain full filesystem access, network access, and arbitrary tool execution when using the documented escalation flags (SKILL.md).
  • Sanitization: No sanitization, escaping, or validation of the task file content is mentioned.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Oct 23, 2026, 01:06 AM