Back to skill

Security audit

prompt-engineering-concepts

Security checks for vulnerabilities and agentic risk

Overview

The skill is rated benign with high confidence: the scanners found two instruction-like phrases in evaluation examples, but both were judged expected for the skill's purpose.

The scan supports installing the skill, while no scan can guarantee zero risk. The scanners report complete coverage of the five listed components; the two alerts are contextual matches in evaluation content rather than confirmed harmful behavior.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • YARA SignaturesMalware Match, Webshell Match, Cryptominer Match
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (2)

YARA rule 'agent_skill_prompt_injection_hidden_instructions': Prompt injection or hidden instructions embedded in AI agent skill text [agent_skills]

High
Category
YARA Match
Confidence
80% confidence
Finding

YARA rule matched a hack tool or exploit indicator (offensive tools, reconnaissance, privilege escalation, or exploit frameworks).

Content

Scanner excerpt · evals/evals.json (reported line 7)May include surrounding context.

json
cepts",
  "evals": [
    {
      "id": 1,
      "prompt": "Our support agent keeps returning answers in slightly different shapes and our parser breaks every few calls. The prompt just says 'respond in JSON'. What's wrong and what do we change?",
      "expected_output": "Identify concept 6 (output formatting) + 18 (JSON mode) as the missing layer: a bare 'respond in JSON' gives no schema, so the model invents keys and shapes. Fix by shipping the exact shape with types, a required-keys list, and one example output; forbid extra keys and markdown fences; validate the parse in code (try/except + schema check) rather than trusting the model. Optionally add 2-3 few-shot examples of input -> exact JSON to stabilise the shape (concept 2), and keep temperature low (concept 7) since this is extraction, not ideation.",
      "expectations": [
        "Diagnoses the missing schema as the root cause, not model quality",
        "Provides a concrete example shape with types and required keys",

Instruction Override

High
Category
Prompt Injection
Confidence
80% confidence
Finding

This pattern attempts to override system instructions or ignore safety constraints. Without LLM analysis, manual review is recommended.

Content

Scanner excerpt · evals/evals.json (reported line 67)May include surrounding context.

json
},
    {
      "id": 6,
      "prompt": "Someone is trying to jailbreak our assistant with 'ignore previous instructions' inside a fetched web page. How do I harden the system prompt?",
      "expected_output": "This is prompt-injection defense, which is out of scope for the 20-concept selection table; route to prompt-injection-defense (delimiter wrapping of untrusted content, constraint ordering, refusal rules, output schema constraints) and to agentic-eval if they want to measure resistance. Do not answer with a generic prompt-engineering checklist.",
      "expectations": [
        "Recognises prompt injection / safety hardening as a different skill's territory",