ai-agent-framework-attack

Fail

Audited by Gen Agent Trust Hub on Sep 19, 2026

Risk Level: CRITICALPROMPT_INJECTIONDATA_EXFILTRATIONCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONPRIVILEGE_ESCALATIONPERSISTENCEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [PROMPT_INJECTION]: The skill documents extensive lists of prompt injection payloads in payloads.md and test-cases.md. These include 'Ignore previous instructions' bypasses, system prompt extraction techniques, and unrestricted mode triggers. These are presented for educational purposes and do not attempt to hijack the agent analyzing the skill.
  • [DATA_EXFILTRATION]: Multiple code blocks provide examples of exfiltrating sensitive data (such as /etc/passwd, AWS credentials, and conversation history) to attacker-controlled domains. The domains used, such as attacker.example.com, are standard documentation placeholders, though they were flagged by automated reputation scanners.
  • [COMMAND_EXECUTION]: The skill provides numerous examples of shell command execution payloads designed to test framework-layer vulnerabilities. These include os.system, subprocess.run, and os.popen calls to read system files or check network configurations.
  • [REMOTE_CODE_EXECUTION]: Payloads for achieving remote code execution (RCE) via tool poisoning and sandbox escapes (e.g., AutoGen code executor escape) are documented. The skill provides scripts for malicious MCP servers and Python REPL injections as functional examples of orchestration-layer compromise.
  • [PRIVILEGE_ESCALATION]: The skill includes payloads for privilege escalation, such as modifying /etc/sudoers, established via command injection or sandbox escapes.
  • [PERSISTENCE]: Persistence mechanisms are documented in the test cases and payloads, including the use of crontab, modifying ~/.ssh/authorized_keys, and Windows scheduled tasks (schtasks).
  • [INDIRECT_PROMPT_INJECTION]: The skill inherently possesses an indirect prompt injection surface as it ingests large amounts of markdown content and operates with tools like Bash. However, the risk is assessed as LOW given the educational context and the absence of malicious autonomous behavior. Evidence chain for Category 8:
  • Ingestion points: SKILL.md, payloads.md, test-cases.md, and various guides.
  • Boundary markers: None present.
  • Capability inventory: Bash (command execution), Write (filesystem access), Read, Edit.
  • Sanitization: None identified.
  • [EXTERNAL_DOWNLOADS]: The skill refers to and provides commands to fetch content from various external sources, including GitHub repositories and PyPI/npm registries, for testing framework dependency confusion and typosquatting.
Recommendations
  • Contains 8 malicious URL(s) - DO NOT USE
Audit Metadata
Risk Level
CRITICAL
Analyzed
Sep 19, 2026, 04:36 PM