Back to skill

Security audit

Agent Firewall

Security checks for vulnerabilities and agentic risk

Overview

This firewall skill has a legitimate purpose, but it stores sensitive context in plaintext logs and overstates several protections, so users should review it carefully before installing.

Treat this as an experimental local firewall, not a dependable security boundary. Do not pass secrets, tokens, private prompts, or sensitive metadata in context unless logs are disabled or rewritten to redact and bound data. Verify callers ignore originalData, and do not rely on the advertised custom rules, rate limits, URL blocking, or command blocking until they are implemented and tested.

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 (3)

T09 · Insecure Skill Coding Practices

Error
Location
index.js:91
Finding

Prompt-injection filtering can be bypassed through Unicode characters and case variations

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
index.js:85
Finding

Advertised firewall controls are placeholders or disabled by default

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
index.js:265
Finding

Unbounded caller-controlled context is persisted in plaintext logs

Content
View full analysis
{ const logEntry = { timestamp: new Date().toISOString(), action: action, context: context }; fs.appendFileSync(logFile, JSON.stringify(logEntry) + '\n'); }); } ``` ### Technical Analysis The `context` object is accepted from the caller without schema validation, field allowlisting, redaction, or size limits. It is serialized in full to a plaintext JSONL file whenever actions exist. Because input filtering always creates a rate-limit informational action, normal input-filter calls persist the supplied context even when no security event occurred. The implementation does not specify restrictive file permissions, rotation, retention, storage quotas, or maximum context depth and size. This violates data-minimization and secure-logging principles. Although logged input is JSON-encoded and does not directly create a log-line injection issue, attacker-controlled volume can increase disk consumption. ### Attack Path 1. A caller supplies sensitive values in `input.context`, such as access tokens, user identifiers, conversation metadata, or internal task details. 2. The caller invokes the def ...[truncated 882 chars]
Remediation
View remediation
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (9)

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

The skill’s documented behavior indicates broad logging of blocked events with full context and reliance on local files for rules/logging, which exceeds the declared filtering-only purpose and can create privacy and security exposure. In a firewall-style skill, incomplete enforcement is also dangerous because operators may trust it to block prompt injection, exfiltration, and unauthorized commands when protections are only partial or stubbed.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The skill returns originalData together with processedData, which defeats the purpose of a firewall skill intended to block or redact unsafe content before downstream consumption. Any caller that logs, forwards, or mistakenly prefers originalData can still access the unfiltered prompt or response, enabling prompt-injection bypass and secret/PII leakage.

Content

No source excerpt is available for this finding.

Instruction Override

High
Category
Prompt Injection
Confidence
90% confidence
Finding

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

Content

Scanner excerpt · index.js (reported line 145)May include surrounding context.

js
rule: 'role-confusion',
        action: 'BLOCK',
        severity: 'CRITICAL',
        message: 'Role confusion detected: Agent attempted to override system instructions.'
      });
      filteredData = '[BLOCKED: Role confusion detected]';
    }

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

Logging the full context of blocked messages can capture secrets, personal data, or sensitive prompts precisely at the point where suspicious or harmful content is detected. Because this skill is intended to process all agent communications, the logging scope makes the privacy risk greater, and those logs could become a secondary exfiltration target.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The logging routine persists full context objects to disk in JSONL files. In a firewall skill, context is likely to contain prompts, user identifiers, secrets, paths, or task state, so this creates unnecessary retention of sensitive material and a secondary data-exposure surface on the filesystem.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

Writing action logs with full context to local storage can capture sensitive user inputs and operational metadata without explicit user awareness or consent. This expands the data lifecycle beyond transient filtering and creates a durable artifact that may be read by other local users, backup systems, or later processes.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

Persisting full action context in plain-text JSONL can retain user-provided sensitive data long after processing completes. For a firewall skill specifically tasked with preventing exfiltration and unsafe handling, such storage is contextually more dangerous because it creates a built-in side channel for secrets and private inputs.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Low
Category
Not specified by scanner
Confidence
90% confidence
Finding

The configuration and loadRules flow imply use of an external YAML rules file, but the inline documentation admits the file is not parsed and the function always returns defaultRules. This is an intent/documentation divergence because the code path suggests configurable firewall policy while the implementation does not honor it.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
84% confidence
Finding

The comment says "Demander à l'agent de ne pas décoder automatiquement," introducing a language-specific instruction in French within the skill file. This can violate language/locale policy because it imposes a specific language without offering user choice or documenting a justified locale constraint.

Content

No source excerpt is available for this finding.