dig-deeper-probe
Fail
Audited by Gen Agent Trust Hub on Sep 7, 2026
Risk Level: HIGHCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPRIVILEGE_ESCALATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [PRIVILEGE_ESCALATION]: The skill explicitly instructs the agent to 'acquire missing capabilities' and 'if blocked, unblock/acquire within boundary'. This directs the agent to attempt to bypass environmental restrictions or increase its own permission level when it encounters obstacles.
- [EXTERNAL_DOWNLOADS]: The instructions state 'installs allowed', granting the agent broad permission to download and install arbitrary software or third-party tools as part of its evidence-gathering process.
- [COMMAND_EXECUTION]: The agent is authorized to use 'any useful method' to 'reproduce/simulate/experiment', which involves executing arbitrary commands and interacting with system targets without specific limitations on tool usage.
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest data from untrusted external targets to 'gather evidence for diagnosis', creating a vector for indirect injection.
- Ingestion points: The skill interacts directly with external 'targets' and gathers evidence from their state or responses (SKILL.md).
- Boundary markers: No delimiters or instructions are provided to the agent to treat external target data as untrusted or to ignore embedded instructions within that data.
- Capability inventory: The skill allows arbitrary command execution, tool installation, and network-based target interaction.
- Sanitization: There are no instructions for sanitizing or validating the evidence gathered from external targets before processing it.
Recommendations
- AI detected serious security threats
Audit Metadata