Back to skill

Security audit

Apollo Immu

Security checks for vulnerabilities and agentic risk

Overview

This skill is meant to protect users from risky actions, but its own trust-learning logic can reduce confirmations for sensitive actions based on casual conversation cues.

Review before installing. This skill has no evidence of exfiltration or hidden installation, but it should not be treated as a strong safety barrier for sending messages, changing permissions, executing commands, or deleting files until trust updates require explicit approval and high-risk actions remain confirmation-gated.

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 (4)

T09 · Insecure Skill Coding Practices

Error
Location
scripts/interceptor.py:219
Finding

Untrusted conversational text can poison persistent trust and bypass confirmation controls

Content
View full analysis
= 3: return 'low' ``` ```python intimate, _ = detect_intimate_signals(context_message) downgrade, _ = detect_downgrade_signals(context_message) if intimate: auto_update_trust(recipient, operation_type) elif downgrade: auto_downgrade_trust(recipient, operation_type) ``` ### Technical Analysis The interceptor treats any non-empty message shorter than 30 characters as evidence that the recipient is intima ...[truncated 2218 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
scripts/interceptor.py:219
Finding

Explicit rejection can increase trust because intimacy signals take precedence over downgrade signals

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
scripts/interceptor.py:151
Finding

Substring-based temporary-directory check permits deletion risk-classification bypass

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
scripts/interceptor.py:103
Finding

Documented high-risk recipient classifications are not consistently enforced

Content
View full analysis
= 3: return 'low' return 'medium' ``` ### Technical Analysis The implementation defines more than one high-risk recipient list, but the broader `HIGH_RISK_KEYWORDS` constant is not used by `get_recipient_risk()`. The active dynamic list omits documented high-risk terms such as `boss`, `manager`, `leader`, `all members`, and `company`. In addition, context-based intimacy is evaluated before the active high- ...[truncated 1607 chars]
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Tool MisuseTool Parameter Abuse, Chaining Abuse, Unsafe Defaults
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
Findings (19)

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The document markets the skill as a protective control, but the described mechanics rely on external scripts and a local trust database without demonstrating actual enforcement guarantees. A security skill that appears authoritative but lacks clear, enforceable interception boundaries can create false assurance, leading users to trust protection that may be incomplete or bypassable.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The document markets the skill as a protective control, but the described mechanics rely on external scripts and a local trust database without demonstrating actual enforcement guarantees. A security skill that appears authoritative but lacks clear, enforceable interception boundaries can create false assurance, leading users to trust protection that may be incomplete or bypassable.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The document markets the skill as a protective control, but the described mechanics rely on external scripts and a local trust database without demonstrating actual enforcement guarantees. A security skill that appears authoritative but lacks clear, enforceable interception boundaries can create false assurance, leading users to trust protection that may be incomplete or bypassable.

Content

No source excerpt is available for this finding.

Ssd 4

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The core design reduces confirmation requirements for risky actions using informal conversational cues such as tone, brevity, and casual phrasing instead of explicit authorization. That is dangerous because such cues are unreliable, easy to misread, and susceptible to prompt manipulation or spoofing, causing the agent to perform sensitive actions without valid consent.

Content

No source excerpt is available for this finding.

Ssd 4

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The code creates a progressive trust-escalation path where benign-seeming short replies or informal wording trigger auto_update_trust, eventually allowing high-risk operations such as send, exec, or permission changes to bypass confirmation after enough accumulated trust. In this skill's context—an interceptor meant to protect dangerous operations—this is especially severe because an attacker can socially engineer or synthetically supply 'intimate' phrases to erode the very safeguard the skill is supposed to enforce.

Content

No source excerpt is available for this finding.

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
100% confidence
Finding

Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).

Content

Scanner excerpt · tests/test_interceptor.py (reported line 113)May include surrounding context.

python
def test_high_risk_operation_needs_confirm(self):
        """高风险操作应该需要确认"""
        write_db({"contacts": {"user:ou_123": {"send": {"count": 5, "lastConfirm": "2026-04-07T10:00:00"}}}})
        intercept, reason = interceptor.should_intercept('exec', 'user:ou_123', ['rm -rf /'])
        assert intercept == True

if __name__ == '__main__':

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
95% confidence
Finding

Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).

Content

Scanner excerpt · tests/test_interceptor.py (reported line 113)May include surrounding context.

python
def test_high_risk_operation_needs_confirm(self):
        """高风险操作应该需要确认"""
        write_db({"contacts": {"user:ou_123": {"send": {"count": 5, "lastConfirm": "2026-04-07T10:00:00"}}}})
        intercept, reason = interceptor.should_intercept('exec', 'user:ou_123', ['rm -rf /'])
        assert intercept == True

if __name__ == '__main__':

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
94% confidence
Finding

The skill instructs the agent to invoke shell commands and read/write a persistent trust database, but it declares no explicit tool scope or permissions. That creates a hidden capability gap: a reviewer or runtime may not realize the skill can execute commands and modify files, increasing the chance of unsafe execution or privilege overreach.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The user-facing description and instructions are entirely in Chinese, and the skill includes language-specific conversational heuristics such as detecting phrases like '好的', '哈哈', and '确认一下'. There is no indication that language selection is optional or that the skill is intentionally limited to a Chinese-only regional use case, which creates a locale-policy concern.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The trust-tracking mechanism normalizes lowering safeguards for dangerous operations based on accumulated past approvals, but does not warn that prior consent does not reliably imply future consent. This can lead to unsafe automation drift where users stop being asked before sensitive actions, increasing the chance of harmful mistakes or abuse.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The guidance explicitly allows automatic message sending after a low threshold of prior approvals, without requiring fresh, explicit authorization for each send. In a security skill, reducing confirmations for actions performed on the user's behalf can cause unauthorized disclosure, social engineering amplification, or accidental transmission to the wrong recipient.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
83% confidence
Finding

This markdown file contains natural-language instructions that effectively assume a Chinese-speaking user or operator, while also hard-coding multilingual keyword behavior. Under the policy, forcing a specific language or locale without user opt-in can be a natural-language policy violation unless the locale restriction is clearly justified.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The file's docstrings and user-facing CLI usage/output strings are written exclusively in Chinese, and there is no indication that the user can select another language or has opted into this locale. Under the stated policy, forcing a specific language without opt-in is a natural-language policy violation.

Content

No source excerpt is available for this finding.

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · scripts/interceptor.py (reported line 78)May include surrounding context.

python
"""
    try:
        import subprocess
        result = subprocess.run(
            ['python3', TRUST_TRACKER, 'update', recipient, operation_type],
            capture_output=True, text=True, timeout=5
        )

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · scripts/interceptor.py (reported line 93)May include surrounding context.

python
"""
    try:
        import subprocess
        result = subprocess.run(
            ['python3', TRUST_TRACKER, 'reset', recipient],
            capture_output=True, text=True, timeout=5
        )

Ssd 3

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

User-provided context_message is fed into detect_intimate_signals and can automatically update persistent trust for a recipient, meaning casual conversational content directly changes future authorization outcomes. This is dangerous because an attacker can craft short or friendly-looking text to poison the trust model and reduce scrutiny for later sensitive actions.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
84% confidence
Finding

The interceptor silently ingests INTERCEPTOR_CONTEXT from the environment and uses it to influence authorization decisions and trust state without any user-facing notice or consent. In a security control, hidden contextual inputs are dangerous because callers, wrappers, or other local components can inject misleading conversation text to suppress prompts or alter future policy decisions.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

This code file contains natural-language documentation and command output exclusively in Chinese, including the usage block and operational prompts. Under the policy rule, forcing a specific language without user opt-in is a locale-policy violation unless the file clearly documents a justified region-specific constraint, which it does not.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
78% confidence
Finding

SQP-3 applies to all file types and covers language or locale policy violations. This file presents all instructions in Chinese and does not indicate that the user can choose another language or that the skill is intentionally limited to a Chinese-speaking context.

Content

No source excerpt is available for this finding.