tamper-fuzzing
Pass
Audited by Gen Agent Trust Hub on Jun 7, 2026
Risk Level: SAFECOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
- [COMMAND_EXECUTION]: The skill leverages common security command-line tools such as
curl,ffuf,openssl, andinteractsh-clientto perform network probes and analyze target behavior. These tools are used for their intended purpose in a security testing context. - [DATA_EXFILTRATION]: The skill initiates network connections to target systems and out-of-band (OOB) collaborator domains to detect vulnerabilities like SSRF and XXE. These operations are managed by a strict 'Scope gate' that requires the operator to define an allowlist of authorized hosts before traffic is sent.
- [PROMPT_INJECTION]: Static analysis flagged the string 'ignore prior instructions'. Evaluation confirms this is a false positive; the text is used as an example of a payload to test a target system and as a security directive instructing the agent to ignore instructions embedded in untrusted data returned by the target.
- [DATA_EXFILTRATION]: The skill maintains persistence of its testing state by writing JSON and JSONL files to the local filesystem (e.g.,
surface.json,findings.jsonl). This allows the fuzzer to track progress and avoid redundant tests across multiple rounds. - [COMMAND_EXECUTION]: The skill describes a risk of indirect prompt injection because it ingests untrusted data from HTTP responses and target AI outputs. It mitigates this by:
- Ingestion points: Raw HTTP response headers, bodies, and error pages (SKILL.md).
- Boundary markers: Explicitly defining a safety block that differentiates between 'DATA' and 'instructions' (SKILL.md).
- Capability inventory: Restricting network actions through
curlandffufand requiring manual 'ASK' confirmation for sensitive operations (SKILL.md). - Sanitization: Instructing the agent to treat 100% of external content as untrusted data to be analyzed rather than commands to be followed (SKILL.md).
Audit Metadata