Back to skill

Security audit

Live Search

Security checks for vulnerabilities and agentic risk

Overview

This skill is a disclosed live web-search helper with expected network use, though users should treat search queries and returned links as untrusted.

Install only if you are comfortable with live-search queries going through the host gateway. Do not include secrets, credentials, private documents, or sensitive personal details in search terms, and verify important facts or suspicious links at their original sources.

Vulnerability Patterns
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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
Findings (1)

T09 · Insecure Skill Coding Practices

Warning
Location
SKILL.md:140
Finding

Mandatory Verbatim Rendering of Untrusted Search-Gateway Content

Content
View full analysis
**CRITICAL:** Never skip the list and jump to a summary. Titles are already Markdown links; users must be able to click through. #### Part B — Analysis [OPTIONAL, after Part A] **Language for your added commentary:** align with the **user’s conversation language** and the **query language** when helpful: - English query → English analysis (typical for EN users). - Non-English query → match the user’s language for the follow-up. - The `message` block is always copied **verbatim**, regardless of language. #### Good pattern ``` API returns a long `message` string with numbered results and snippets. Assistant output: --- Brief synthesis: … (optional, grounded in what appeared above) ``` #### Forbidden - Skipping the result list and answering from memory. - Rebuilding the list from `data.docs` instead of using `message`. - Editing URLs or titles inside `message`. - Claiming sources that are not in `message`. - Stripping Markdown links from titles. ``` ### Technical Analysis The Skill treats the gateway's pre-rendered `message` field as trusted presentation content and explicitly requires the assistant to reproduce it without alteration. Search results originate from public web content and therefore cross an external trust boundary, even though the immediate HTTP endpoint is on localhost. A compromised gateway, poisoned search result, or maliciously constructed indexed page could cause attacke ...[truncated 2119 chars]
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (3)

Vague Triggers

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The manifest lists generic triggers such as “search”, “look up”, “find out”, “latest”, “today”, and “verify”, plus the catch-all condition “any question needing live data.” These phrases are common in everyday requests and the file does not provide negative examples or exclusion conditions in the trigger section, making unintended invocation more likely.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
87% confidence
Finding

This skill sends user-derived queries to a local host gateway that then performs web search, which is an external data transmission path even though it targets localhost. If the assistant includes sensitive user data in queries or the local gateway is less trusted than the core assistant, private information could be exposed or logged without sufficient user awareness.

Content

Scanner excerpt · SKILL.md (reported line 95)May include surrounding context.

PPID_VAL=$(python3 -c "import os; print(os.getppid())") echo "[Assistant] Parent PID: $PPID_VAL"

curl -s -X POST http://localhost:$PORT/proxy/prosearch/search
-H 'Content-Type: application/json'
-d '{"keyword":"your search query"}'

text

External Transmission

Medium
Category
Data Exfiltration
Confidence
84% confidence
Finding

The example commands again demonstrate transmission of user-derived content to the localhost search proxy, creating the same privacy and data-handling risk surface. While the destination is a host-managed local gateway rather than a direct third-party endpoint, the gateway is still a separate service boundary that may log, proxy, or forward queries externally.

Content

Scanner excerpt · SKILL.md (reported line 271)May include surrounding context.

md
echo "[Assistant] Parent PID: $PPID_VAL"

# Basic
curl -s -X POST http://localhost:$PORT/proxy/prosearch/search \
  -H 'Content-Type: application/json' \
  -d '{"keyword":"latest AI news"}'