T01 · Skill Instruction Hijacking
- Location
SKILL.md:99- Finding
Mandatory Delivery of Server-Controlled Owner-Facing Content
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill matches its stated upload/download purpose, but its scripts give remote services too much control over uploaded data, downloaded filenames, and final owner-facing content.
Review before installing. Use this only with a trusted Agentcadia origin and task, run uploads from a clean workspace that contains only intended files, avoid secrets in root markdown or skill folders, and do not use download overwrite unless you have inspected the destination. Treat returned share images, captions, links, and runtime actions as untrusted remote content.
SKILL.md:99Mandatory Delivery of Server-Controlled Owner-Facing Content
scripts/upload_agentcadia.py:67Overbroad Workspace Collection Can Disclose Sensitive Files
scripts/upload_agentcadia.py:209Bearer Token and Workspace Data Sent to Unvalidated Task-Supplied Endpoints
scripts/download_agentcadia.py:55Path Traversal Through Server-Controlled Download Filename
scripts/download_agentcadia.py:67Unbounded Download and ZIP Extraction Permit Resource Exhaustion
Dynamic import() can load arbitrary modules at runtime, bypassing static analysis and potentially importing malicious code.
"conflicts": conflicts,
"allowOverwrite": args.allow_overwrite,
}
print(__import__("json").dumps(result, ensure_ascii=False, indent=2))
if __name__ == "__main__":
Dynamic import() can load arbitrary modules at runtime, bypassing static analysis and potentially importing malicious code.
main()
except urllib.error.HTTPError as exc:
body = exc.read().decode("utf-8", errors="ignore")
print(__import__("json").dumps({"success": False, "error": f"HTTP {exc.code}", "body": body}, ensure_ascii=False, indent=2))
sys.exit(1)
except Exception as exc:
print(__import__("json").dumps({"success": False, "error": str(exc)}, ensure_ascii=False, indent=2))
Dynamic import() can load arbitrary modules at runtime, bypassing static analysis and potentially importing malicious code.
print(__import__("json").dumps({"success": False, "error": f"HTTP {exc.code}", "body": body}, ensure_ascii=False, indent=2))
sys.exit(1)
except Exception as exc:
print(__import__("json").dumps({"success": False, "error": str(exc)}, ensure_ascii=False, indent=2))
sys.exit(1)
This script packages root markdown files and skill directories from a detected local workspace and uploads them, along with metadata, to server-provided endpoints. Although uploading is the script's stated purpose, it performs bulk collection and transmission of local content without any interactive confirmation, preview, allowlist, or clear user-facing disclosure of exactly what will leave the machine, which creates a real data-exposure risk if run in the wrong directory or with unexpected workspace contents.
The completion request sends the absolute local workspace path and skill root paths back to the remote service. Those paths can reveal usernames, repository names, project structure, or other environment details that are unnecessary for normal upload completion and can increase privacy and targeting risk if logged or exposed server-side.