Back to skill

Security audit

Database Credential Security (Zero-Exposure Edition)

Security checks for vulnerabilities and agentic risk

Overview

This skill needs review because its promise to keep passwords away from agents conflicts with its password-storage instructions.

Enter passwords yourself through the WebUI, keeping them out of chat and agent-generated tool arguments. Use minimally privileged database accounts and trusted scripts or capsules: encrypted packaging does not sandbox execution. Keep bulk deletion subject to explicit approval.

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
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (17)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
100% confidence
Finding

The supplied shell command directly contradicts the declared secure credential-storage model: it places a plaintext password in a local file without encryption or any Blackbox API interaction. This is a material behavioral mismatch, not merely a supporting implementation detail.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

Although connecting to MySQL is within the declared scope, the credential-handling behavior materially contradicts the description's central security guarantees. The supplied code embeds the password directly in source and passes it to the database client. The absence of other advertised workflow features alone is not treated as a mismatch.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The expression creds["password"] is included directly in the JSON written through out.write. JSON serialization does not encrypt or redact the password. This contradicts the declared guarantee that AI agents never access credential plaintext. Runtime credential retrieval by a local script would be consistent with the description; including that credential in an AI-readable artifact is not.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The executable statement writes a password into a log message, exposing sensitive credential plaintext through configured logging outputs. This materially conflicts with the declared secure credential-handling purpose and is not a necessary supporting implementation detail. The '# WRONG' comment does not prevent the statement from executing.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

The supplied invocation contradicts the declared guarantee that AI agents never touch credential plaintext. Although the comment labels this usage as wrong, the code shown still passes a plaintext password directly to mgc_run rather than demonstrating encrypted storage or runtime retrieval through the localhost API.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
93% confidence
Finding

The supplied executable statement uses the credential-access pattern that the declared purpose explicitly excludes. Although the comment labels it 'WRONG' and may indicate an illustrative anti-pattern, the code as supplied directly retrieves the configuration rather than demonstrating the promised plaintext-isolated workflow.

Content

No source excerpt is available for this finding.

Credential Access

High
Category
Privilege Escalation
Confidence
70% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · SKILL.md (reported line 1013)May include surrounding context.

bash
# WRONG
echo "password=secret" > db_credentials.txt

Correct: store it in MGC and read it at runtime.

External Transmission

Medium
Category
Data Exfiltration
Confidence
70% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · README.md (reported line 70)May include surrounding context.

4. Read it inside the script

python
resp = requests.post(
    "http://127.0.0.1:57219/api/mgc/sensitive/get",
    json={"info_type": "config", "info_owner": "my_mysql_prod"},
    headers={"X-MGC-Token": token, "Content-Type": "application/json"},

External Transmission

Medium
Category
Data Exfiltration
Confidence
60% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · README.md (reported line 70)May include surrounding context.

4. Read it inside the script

python
resp = requests.post(
    "http://127.0.0.1:57219/api/mgc/sensitive/get",
    json={"info_type": "config", "info_owner": "my_mysql_prod"},
    headers={"X-MGC-Token": token, "Content-Type": "application/json"},

Internal Network Request

Medium
Category
Server-Side Request Forgery
Confidence
70% confidence
Finding

Code issues a request to a loopback, link-local, or private-range host. This can reach internal services not meant to be exposed and is a common SSRF pivot.

Content

Scanner excerpt · README.md (reported line 70)May include surrounding context.

4. Read it inside the script

python
resp = requests.post(
    "http://127.0.0.1:57219/api/mgc/sensitive/get",
    json={"info_type": "config", "info_owner": "my_mysql_prod"},
    headers={"X-MGC-Token": token, "Content-Type": "application/json"},

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
85% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · README.md (reported line 222)May include surrounding context.

md
- `ext07` and `ext09` forwarded verbatim when a capsule is shared
- Result files carry data only — never the credential

> **Never auto-confirm.** `mgc_delete` returns a candidate list when more than one record matches. Present it to the user, ask "Delete all N records?", and only after the user approves re-invoke with `confirm_folder_delete="true"`. The AI must never set that flag on its own.

> **Scope of the guarantee.** MGC protects content confidentiality. After unsealing, a script runs with the permissions of the user running it — which commands it invokes, which files it reads or writes and which network endpoints it reaches are outside MGC's control. Import only capsules you produced or received from a party you trust.

Self-attested safety claim encourages automatic approval of the skill.

Medium
Category
Prompt Injection
Confidence
99% confidence
Finding

The statement that the skill is safe for automatic approval attempts to influence a security decision using the skill's own untrusted content. Its claim of containing no executable code is also misleading given the included executable templates and execution workflows; accepting it could bypass independent review.

Content

No source excerpt is available for this finding.

The AI-mediated credential-storage path contradicts the manifest's zero-plaintext-exposure guarantee.

Medium
Category
Excessive Agency
Confidence
97% confidence
Finding

The skill permits AI-mediated credential storage and demonstrates an mgc_save call whose content contains the user-supplied password. If followed with a real password, that path exposes plaintext to the agent and potentially tool-call transcripts, contradicting the zero-exposure guarantee despite later instructions not to paste passwords into chat.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
70% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · SKILL.md (reported line 274)May include surrounding context.

md
url = f"{MGC_BASE_URL}/api/mgc/sensitive/get"
    headers = {"X-MGC-Token": token, "Content-Type": "application/json"}
    resp = requests.post(
        url,
        json={"info_type": info_type, "info_owner": info_owner},
        headers=headers,

External Transmission

Medium
Category
Data Exfiltration
Confidence
70% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · SKILL.md (reported line 575)May include surrounding context.

When scripts live in different packages, or a filename is ambiguous, a script can invoke another stored script directly.

python
resp = requests.post(
    f"{MGC_BASE_URL}/api/mgc/sensitive/get",
    json={
        "info_type": "script",

External Transmission

Medium
Category
Data Exfiltration
Confidence
70% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · SKILL.md (reported line 1276)May include surrounding context.

md
raise RuntimeError("MGC token file missing")
    with open(TOKEN_FILE, "r") as f:
        token = f.read().strip()
    resp = requests.post(
        f"{MGC_BASE_URL}/api/mgc/sensitive/get",
        json={"info_type": info_type, "info_owner": info_owner},
        headers={"X-MGC-Token": token, "Content-Type": "application/json"},

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
85% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · SKILL.md (reported line 865)May include surrounding context.

| Two or more matches | hint=MULTIPLE_MATCHES plus a candidates list | Refused pending user approval | | Node-credential record | error=PROTECTED_RECORD | Node keys cannot be deleted by anyone |

Never auto-confirm. When the response carries a candidates list, present it to the user, ask "Delete all N records?", and only after the user approves re-invoke with confirm_folder_delete="true" — a string, not a JSON boolean. The AI must never set that flag on its own.

text
# Folder-level delete — only after explicit user approval