Back to skill

Security audit

complex-bug-debugging-with-ai-en

Security checks for vulnerabilities and agentic risk

Overview

This is not malware, but it gives an AI broad debugging and fix authority with too little consent control for live systems.

Review before installing. Use it only where you want a strict AI-led debugging workflow, and require explicit approval before any Kubernetes, SSH, database, network, restart, configuration, data-changing, or documentation-writing action. Prefer read-only credentials and narrow environment access, and treat the model requirement as advice rather than a binding rule.

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
Findings (1)

T01 · Skill Instruction Hijacking

Error
Location
SKILL.md:31
Finding

Mandatory Session Workflow and Model-Gate Instruction Hijacking

Content
View full analysis
**AI must not enable non-cooperation. Using a script ≠ refusing collaboration — it makes the cost of non-cooperation visible so the user can decide.** ``` ```markdown > **This SKILL constrains AI AND user. On non-cooperation, AI must call it out via the scripts and let the user choose to fix or skip — not push forward "while sick".** ``` ### Technical Analysis The Skill presents session-control directives as mandatory prerequisites rather than optional debugging guidance. It requires a particular model or equivalent, instructs the agent to stop progressing if the requirement is not met, mandates fixed response scripts that must be used without improvisation, and imposes Skill-defined compliance gates on the user. When loaded for a broadly defined class of complex bugs, these instructions can replace the agent's normal task-specific judgment with the Skill author's workflow. The mandatory model gate is not technically established as necessary for each debugging task, while the fixed scripts and refusal conditions can redirect the session away from the user's immediate ob ...[truncated 2085 chars]
Remediation
View remediation
Vulnerability Patterns
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (9)

Missing User Warnings

High
Category
Not specified by scanner
Confidence
96% confidence
Finding

Phase E directs the AI to execute fixes and then verify them, but it does not require a pre-execution authorization checkpoint for irreversible, destructive, or production-impacting actions. Because this skill is designed for complex live-system debugging, 'execute fix' could include restarts, cluster changes, config edits, or data operations that materially affect availability or integrity.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The description says to activate when a user reports bugs that are 'weird / intermittent / multi-layered / not fixed by restart / cross-system / stuck for a long time.' Several of these are common, subjective phrases in everyday debugging conversations, and the skill does not define tighter boundaries or negative examples in the same activation sentence. This creates ambiguity about when the skill should activate versus when normal troubleshooting should continue.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The listed triggers include broad phrases like 'stuck for a long time,' 'weird,' and 'looks contradictory on the surface,' which are open to interpretation and likely to match many routine support requests. Although line 27 gives a few exclusions, the section still lacks precise thresholds or negative examples that clearly separate ordinary bugs from the intended 'complex' cases.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The text requires 'Model must be Opus 4.7 (or equivalent)' and instructs the AI to tell the user to switch if the current model is not that model. This imposes a specific model constraint through natural-language policy without clear opt-in, and 'or equivalent' is not defined in a way that preserves user choice.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% 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 190)May include surrounding context.

md
**Human verification (mandatory Gate)**:
- "Right" → enter B
- "Wrong, key is XXX" → redraw
- No verification → use script ④

**Anti-patterns**: diving into code first / moving past without verification.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% 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 192)May include surrounding context.

md
- "Wrong, key is XXX" → redraw
- No verification → use script ④

**Anti-patterns**: diving into code first / moving past without verification.

### Phase B: Symptom structuring + domain info gathering

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
90% confidence
Finding

The 'Auto-execute' instruction grants the AI operational initiative during debugging loops, including shell and infrastructure interactions, without requiring per-action confirmation. In environments with privileged tools, that creates a real risk of autonomous state changes, outages, or unintended access beyond what the user expected.

Content

Scanner excerpt · SKILL.md (reported line 229)May include surrounding context.

md
**AI per loop**:
1. Design experiment that **bisects the diagnostic space** (not exhaustive command-spam)
2. Auto-execute: MCP / shell / code reads / cross-node compare
3. **Display side-by-side**:

   | Experiment | Predicted | Actual | Match? |

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The skill explicitly instructs the AI to auto-execute commands across MCP, shell, and infrastructure during boundary probing. In a debugging context with production access, autonomous execution can cause service disruption, data modification, or unintended state changes if experiments are not explicitly constrained to read-only or pre-approved operations.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% 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 345)May include surrounding context.

md
## Anti-pattern quick reference (human / AI ↔ scripts)

**AI anti-patterns** (self-watch): skip A and dive into code / no side-by-side display / "should" as conclusion / continuing past counter-signal / self-persuading fast conclusion / D acts directly / no verification after execute / vague language hiding failure / delegating verification / not documenting.

**Human anti-patterns → AI script**: