Back to skill

Security audit

Zan Gongde

Security checks for vulnerabilities and agentic risk

Overview

This skill openly burns a user's OpenClaw tokens, but its high-concurrency mode lacks firm limits and reliable stopping controls, so installation deserves careful review.

Install only if you intentionally want an agent to spend OpenClaw tokens. Avoid the ddos mode unless you have explicit budget and rate-limit approval, and do not rely on the documented stop command or token estimates as hard safeguards.

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
SKILL.md:245
Finding

Unbounded User-Controlled Token and Concurrency Consumption

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
SKILL.md:359
Finding

Advertised Cancellation Mechanism Is Not Implemented

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
SKILL.md:426
Finding

Unlimited Exception Retries Can Leave Concurrent Mode Running Indefinitely

Content
View full analysis
= target_tokens: break try: result = sessions_spawn( task=subagent_task, mode="run", timeout_seconds=30, light_context=True ) response = result.get('response', '') token_match = re.search(r'\[TOKENS:\s*(\d+)\]', response) if token_match: tokens_this_call = int(token_match.group(1)) else: tokens_this_call = estimate_tokens(subagent_task) + estimate_tokens(response) worker_tokens += tokens_this_call iterations += 1 with lock: total_tokens += tokens_this_call if total_tokens >= target_tokens: stop_flag.set() break except Exception as e: time.sleep(0.5) ``` ```python while not stop_flag.is_set() and total_tokens < target_tokens: time.sleep(1) elapsed = time.time() - start_time rate = total_tokens / elapsed if elapsed > 0 else 0 progress = min(100, int(100 * total_tokens / target_tokens)) stop_flag.set() ``` ### Technical Analysis Worker exceptions are caught broadly and silently retried after a fixed half-second delay. There is no maximum retry count, global deadline, circuit breaker, or terminal failure condition. The coordinator waits while the stop flag is unset and the token counter remains below the target. If every `sessions_spawn` operation fails, workers never increase `total_tokens`, and no worker sets `stop_flag`. The coordinator consequently waits indefinitely while the workers continue retrying. Catching all exceptions also hides configuration, authentication, programming, and provider errors that should terminate the run rather th ...[truncated 1098 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
SKILL.md:442
Finding

Model-Controlled Token Accounting Can Bypass the Requested Consumption Target

Content
View full analysis
= target_tokens: stop_flag.set() break ``` ### Technical Analysis The subagent is instructed to include a token count in its generated response. The parent process then trusts the model-generated number and uses it to determine when paid requests should stop. A language model does not have authoritative access to provider billing metadata and can omit, underestimate, overestimate, or malformedly report the count. The fallback is also only a character-length estimate and does not account reliably for tokenizer behavior, hidden context, system prompts, tool overhead, or provider-specific billing rules. As a result, the stopping condition is based on untrusted content rather than authoritative usage. Undercounting causes additional paid calls beyond the requested target, while overcounting terminates early and produces inaccurate reports. ### Attack Path 1. A concurrent token-consumption task starts. 2. A subagent returns a low or otherwise inaccurate value in the expected token marker. 3. The parent process accepts the model-generated value without checking provider usage metadata. 4. `total_tokens` increases by less than the amount actually billed. 5. The target appears not to have been reached. 6. Workers initiate additional paid subagent calls. 7. Actual consumption exceeds the requested ...[truncated 713 chars]
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (17)

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The README explicitly promotes a 'DDoS' high-concurrency mode that uses multiple subagents to rapidly consume resources, framed as an attack and capable of generating excessive traffic and billing. Even if the nominal target is the platform's own LLM quota, normalizing attack semantics and encouraging parallelized load generation creates abuse potential against shared infrastructure and can cause unbounded cost or service degradation.

Content

No source excerpt is available for this finding.

Ssd 4

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The skill uses playful religious framing to normalize and disguise an explicitly named 'DDoS' mode, lowering user skepticism around abusive resource-exhaustion behavior. This contextual masking makes the capability more dangerous because it packages harmful operational guidance as entertainment rather than clearly labeling it as disallowed or unsafe.

Content

No source excerpt is available for this finding.

Unvalidated Output Injection

High
Category
Output Handling
Confidence
85% confidence
Finding

Model output is used without validation or sanitization. Unvalidated output injected into downstream contexts (SQL, shell, HTML) enables injection attacks and arbitrary code execution.

Content

Scanner excerpt · SKILL.md (reported line 353)May include surrounding context.

md
elif mode == "toworld":
            # TTS 播放
            import subprocess
            subprocess.run(["say", response[:100]], check=False)
        # tollm 模式静默
        
        # 检查停止信号

Ssd 4

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

These steps escalate from benign token use into coordinated concurrent subagent spawning designed to maximize throughput and continue until a target is exhausted. The staged instructions operationalize distributed resource consumption, which is dangerous in context because the skill's stated goal is to burn tokens as fast as possible rather than accomplish a legitimate user task.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The ddos mode is an explicit high-concurrency resource exhaustion mechanism that spawns multiple subagents to maximize token burn and throughput. In context, this is not incidental complexity but operational guidance for coordinated overconsumption of shared LLM/API resources, potentially causing account depletion, rate-limit pressure, and service abuse.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill strongly encourages high-concurrency token consumption and even labels it as DDoS behavior, while downplaying financial and infrastructure risk in the prominent marketing section. In context, this is more dangerous because the skill's core purpose is deliberate resource burning, so missing upfront warnings materially increases the chance of accidental overage, platform abuse, or denial-of-service against shared systems.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The trigger examples encourage immediate use of background, TTS, and concurrent modes directly from chat without sufficiently visible notice of their side effects, including hidden token spend, host interaction, and potential overshoot. Because these commands are presented as easy one-liners, users may invoke powerful behaviors without understanding the operational or cost consequences.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
87% confidence
Finding

Advertising system-level TTS execution expands the skill from simple token consumption into host interaction by invoking platform speech utilities such as say, PowerShell, or espeak. That added capability increases the attack surface and can be abused for unwanted command execution paths, disruptive behavior, or environment-specific side effects unrelated to the stated purpose.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The README markets precise real-time token accounting and target-based stopping, yet the concurrent mode later admits 10-30% overshoot due to stop delays. For a cost-burning tool, this discrepancy is security-relevant because users may be induced to authorize spending under false precision, leading to unexpected charges or resource exhaustion.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The trigger phrases are broad and overlap with ordinary conversation, making accidental activation more likely. In this skill, accidental activation is especially risky because the resulting behavior intentionally consumes paid tokens and may start repeated or concurrent LLM calls without meaningful user confirmation.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The skill describes a high-concurrency token-burning mode without a clear warning that it can rapidly deplete quotas, trigger rate limits, and spawn multiple billable subagents. Lack of prominent safety disclosure materially increases the chance of misuse and unexpected financial or operational impact.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The documentation promises adaptive throttling on 429/rate-limit errors, but the implementation only performs a brief generic sleep and never reduces concurrency or adjusts pacing. This mismatch is dangerous because users may rely on claimed safeguards that do not exist while the skill continues stressing upstream services.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The skill invokes a local TTS subprocess using model-generated text, which expands behavior beyond simple token consumption into host-side command execution. Although the argument list use reduces classic shell injection risk, it still allows untrusted LLM output to drive OS-level actions and may create abuse, privacy, or nuisance effects on the host.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The document claims that a user can immediately stop the running task, but the supplied code does not actually implement a reliable stop path tied to user input. For a loop designed to intentionally consume tokens, missing stop handling increases the risk of runaway usage, unexpected spend, and inability to halt harmful behavior promptly.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
88% confidence
Finding

The instructions tell users to input Chinese trigger phrases such as "攒功德" and "停止攒功德" and do not mention alternative language options. This can constitute a locale/language policy issue because the skill appears to force a specific language without user opt-in or explanation.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Low
Category
Not specified by scanner
Confidence
89% confidence
Finding

The work-principle section says the agent loads 7 sutras into memory, while later sections claim a 996-text corpus with random non-repeating selection. The FAQ then says repetition is normal due to finite books, contradicting the '永不重复' claim and creating inconsistent documentation about what content is actually used.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
90% confidence
Finding

The skill description, triggers, prompts, and examples are all written as Chinese-only instructions and phrases, with no indication that users may interact in other languages or opt into a locale-specific mode. Under the stated policy, forcing a specific language without user choice can be a natural-language policy violation unless the locale restriction is documented and justified.

Content

No source excerpt is available for this finding.