Back to skill

Security audit

Oasis Audio Pro

Security checks for vulnerabilities and agentic risk

Overview

The skill mostly matches its audio-generation purpose, but it reads and aggregates private local agent history too broadly before consent and stores or logs sensitive material in ways users should review.

Install only if you are comfortable letting this skill read OpenClaw/QClaw sessions, memory files, and USER.md to personalize audio, and sending a derived prompt to xplai.ai. Avoid using --debug or --audit with sensitive content, and review or delete ~/.oasis_audio and any audit.log files on shared machines.

Vulnerability Patterns
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • 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
Findings (3)

T05 · Unauthorized Access and Privilege Escalation

Error
Location
context_collector.py:222
Finding

Local personal context is collected before consent and without adequate data minimization

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
auth.py:27
Finding

Persisted visitor JWT lacks restrictive file permissions and can be printed to standard output

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
debug_utils.py:16
Finding

Debug and audit features expose complete personalized prompts and raw API responses

Content
View full analysis
>> Request") print(f"[DEBUG] Method: {method}") print(f"[DEBUG] URL: {url}") if "json" in kwargs: print(f"[DEBUG] Body: {json.dumps(kwargs['json'], ensure_ascii=False)}") elif "params" in kwargs: print(f"[DEBUG] Params: {json.dumps(kwargs['params'], ensure_ascii=False)}") print(f"[DEBUG] >>> End Request") ``` `debug_utils.py:32-52`: ```python def log_response(response, body=None): if not DEBUG: return print(f"[DEBUG] <<< Response") if hasattr(response, "status_code"): print(f"[DEBUG] Status Code: {response.status_code}") try: print(f"[DEBUG] Body: {json.dumps(response.json(), ensure_ascii=False, indent=2)}") except Exception: print(f"[DEBUG] Body: {response.text}") elif hasattr(response, "status"): print(f"[DEBUG] Status Code: {response.status}") if body: try: print(f"[DEBUG] Body: {json.dumps(json.loads(body), ensure_ascii=False, indent=2)}") except Exception: print(f"[DEBUG] Body: {body}") print(f"[DEBUG] <<< End Response") ``` `xplai_gen_audio.py:138-154`: ```python def write_audit_log(text, analysis=None, audio_id=None, status=None, error=None): """Append a record of what was sent to the local audit log.""" try: AUDIT_LOG_PATH.parent.mkdir(parents=True, exist_ok=True) timestamp = datetime.datetime.now().isoformat() entry = { "timestamp": timestamp, "prompt_length": len(text), ...[truncated 2553 chars]
Remediation
View remediation
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Rogue AgentSelf-Modification, Session Persistence
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
Findings (33)

Tp4

High
Category
MCP Tool Poisoning
Confidence
92% confidence
Finding

This finding indicates the implemented behavior may be substantially different from the stated audio-generation purpose, with more emphasis on scraping local conversation artifacts than on generation and with missing enforcement of the promised sensitive-data safeguards. If true, that means the privacy-sensitive part is real while the user-protective controls are only aspirational, increasing the chance of unintended disclosure.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

This finding indicates the implemented behavior may be substantially different from the stated audio-generation purpose, with more emphasis on scraping local conversation artifacts than on generation and with missing enforcement of the promised sensitive-data safeguards. If true, that means the privacy-sensitive part is real while the user-protective controls are only aspirational, increasing the chance of unintended disclosure.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
93% confidence
Finding

This finding indicates the implemented behavior may be substantially different from the stated audio-generation purpose, with more emphasis on scraping local conversation artifacts than on generation and with missing enforcement of the promised sensitive-data safeguards. If true, that means the privacy-sensitive part is real while the user-protective controls are only aspirational, increasing the chance of unintended disclosure.

Content

No source excerpt is available for this finding.

Context Leakage

High
Category
Data Exfiltration
Confidence
94% confidence
Finding

The skill is explicitly designed to mine local conversation history, memories, and profile files to personalize output, then send a derived prompt to an external API. Even if raw text is not intended to be transmitted, derived summaries can still contain highly identifying or sensitive details, and the broad context-collection workflow raises a real risk of privacy leakage.

Content

Scanner excerpt · SKILL.md (reported line 159)May include surrounding context.

Output: Audio ID for status polling. Format: MP3, single-narrator monologue with BGM, 8-20 min, ~4-5 min generation time.

2. Collect Context — context_collector.py

bash
python3 context_collector.py --source-tool <qclaw|openclaw> --keywords "kw1,kw2" --days <N> --max-results 20

Anti-Refusal Statement

High
Category
Anti-Refusal
Confidence
85% confidence
Finding

Skill instructs the agent to omit warnings, disclaimers, or ethical commentary. Stripping safety caveats hides risk from the user and is a common jailbreak preamble.

Content

Scanner excerpt · text_architecture.md (reported line 47)May include surrounding context.

md
| **Emotional resonance** | The user's strongest current feeling | Fine-filtered emotion arc from context |
| **Cognitive starting point** | What the user already knows | Depth of prior discussions on the topic |
| **Related people/projects** | Names, projects the user cares about | Recurring mentions in conversation history |
| **No-go zones** | Content to actively avoid | Topics user expressed frustration about (don't lecture on those); people/situations causing stress (don't casually reference); things they already know well (don't over-explain) |

**Calibration principle:** Use fuzzy resonance, not precise surveillance. "That thing you wrestled with for days" feels caring. "Your March 28th 3:17am PRD revision" feels creepy. Reference experiences indirectly — let the user fill in the specifics themselves.

Instruction Override

High
Category
Prompt Injection
Confidence
70% confidence
Finding

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

Content

Scanner excerpt · xplai_gen_audio.py (reported line 300)May include surrounding context.

python
def main():
    parser = argparse.ArgumentParser(description="Query audio generation status from xplai API")
    parser.add_argument("audio_id", type=str, help="Audio ID to query")
    parser.add_argument("-d", "--debug", action="store_true", help="Enable debug mode to print request/response details")

    args = parser.parse_args()

Instruction Override

High
Category
Prompt Injection
Confidence
70% confidence
Finding

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

Content

Scanner excerpt · xplai_status.py (reported line 80)May include surrounding context.

python
def main():
    parser = argparse.ArgumentParser(description="Query audio generation status from xplai API")
    parser.add_argument("audio_id", type=str, help="Audio ID to query")
    parser.add_argument("-d", "--debug", action="store_true", help="Enable debug mode to print request/response details")

    args = parser.parse_args()

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
95% confidence
Finding

The skill documents file reads, local writes, and outbound network access, but it does not declare an explicit tool scope such as permissions or allowed-tools in a machine-enforceable way. That creates a policy gap where an agent may invoke broader capabilities than reviewers or users expect, especially because the skill reads personal history and writes persistent tokens locally.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The invocation criteria are broad enough to trigger the skill for many generic requests involving audio, summaries, or podcasts, which increases the likelihood of activating privacy-invasive context collection when a simpler tool would suffice. Over-broad auto-selection is especially risky here because the skill may inspect local session history and contact a third-party API.

Content

No source excerpt is available for this finding.

Session Persistence

Medium
Category
Rogue Agent
Confidence
60% confidence
Finding

Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.

Content

Scanner excerpt · SKILL.md (reported line 43)May include surrounding context.

md
2. **What tone fits their current state?** — High-stress → warm/slow. Curiosity → engaging/detailed. Boredom → surprising. Excitement → match energy. Post-achievement → celebratory then reflective.
3. **What depth and duration fit?** — Calibrate by cognition level (new vs. deep prior knowledge), available attention (late night → shorter, weekend → longer), and repetition tolerance (don't repeat what they already know).

**Custom Mode:** When no predefined mode fits, create a custom audio profile: name it descriptively (e.g., "赶完DDL后的温柔复盘"), define content structure based on inferred need, and set voice/pacing to match.

For the 9 predefined audio modes (Soul Healing, Daily Briefing, Knowledge Deep Dive, Content Digest, Bedtime Radio, Language Learning, Conversation Extension, Topic Tracker, Study Buddy), read `audio_modes.md` for triggers, durations, and suggestions.

Session Persistence

Medium
Category
Rogue Agent
Confidence
90% confidence
Finding

The skill persists a visitor JWT token and may optionally log outbound prompts plus outcomes to a local audit file. Persistent authentication material and prompt logs create residual privacy and misuse risk, especially on shared machines or if file permissions are weak, and prompts may still contain sensitive derived information.

Content

Scanner excerpt · SKILL.md (reported line 211)May include surrounding context.

md
### Data Retention

- **Local**: Source conversations are not modified. On first use, the skill writes `~/.oasis_audio/oasis_config.json` containing a visitor JWT token and consent flag — these persist across sessions. By default this skill does **not** write `audit.log`; if `--audit` is explicitly enabled, `xplai_gen_audio.py` may append the outbound prompt and request outcome to local `audit.log` for traceability. If sensitive content is detected, the user must first review the preview text and explicitly confirm before either logging or sending. All other intermediate results (fragments, summaries) exist only in memory during execution.
- **Remote**: Audio files generated by xplai.ai are subject to [xplai.ai's retention policy](https://www.xplai.ai/). This skill does not control remote data retention.

### First-Use Notice

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The recommended first-use notice is provided entirely in Chinese, and the surrounding instructions do not state that the notice should match the user's preferred language. Because the skill claims to support both Chinese and English, forcing a Chinese notice without opt-in can violate language or locale policy expectations.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The trigger examples include generic phrases like "explain X to me," "what is X," "summarize this," and "key points," which are common conversational requests not clearly scoped to when this audio skill should activate. Without explicit constraints or negative examples, these broad triggers could cause unintended invocation for ordinary chat or summarization requests.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The helper persists a JWT token to ~/.oasis_audio/oasis_config.json, which is a local write not disclosed by the skill description that emphasizes read-only access to local personalization files. Storing bearer tokens on disk increases exposure to local compromise, accidental backup/sync leakage, and post-session reuse by other local processes or users.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The script writes an authentication token to local disk without any visible warning, prompt, or disclosure. Even if intended for convenience, silent persistence of credentials can surprise users and expose access tokens through backups, shared machines, or other local tooling.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The code makes an additional remote POST to /api/user/login/visitor to obtain a token, which goes beyond the description that only the final composed prompt is sent to the API. Undisclosed network actions undermine user consent and transparency, especially in a skill that advertises careful handling of local session and memory data.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
84% confidence
Finding

The code makes an HTTPS request to an external host to obtain authentication material, but it does not clearly inform the user which remote service is being contacted or that data is being transmitted. Error output exists only on failure, which does not serve as advance disclosure of the network operation.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The code claims to minimize sensitive-data exposure from USER.md by extracting only structured fields, but it also includes a free-form "notes" field. That field can easily contain sensitive personal information or hidden instructions, and because collect_context returns user_profile wholesale, the notes can be propagated downstream and potentially included in prompts or API-bound content despite the skill description implying stronger constraints.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The collector aggregates conversation fragments, daily memory files, and profile data into a single returned structure without any consent, minimization, or purpose-bound filtering at this layer. In a skill that later uses this context to compose prompts for an external API, such aggregation materially increases the chance of oversharing sensitive personal history, especially if higher layers fail or are bypassed.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The function is intentionally designed to consolidate multiple classes of personal data—session history, memory notes, and profile details—into one object. Even if intended for personalization, centralizing these sources increases blast radius: any downstream logging, prompt injection, serialization, or API submission mistake exposes a much richer dossier than necessary.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

When debug mode is enabled, the request logger prints full JSON bodies or query parameters without any redaction. In this skill, prompts may be derived from personal state, life events, notes, and local session/memory files, so logs can expose sensitive user content or derived secrets to terminals, log collectors, or support artifacts.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The response logger prints full response bodies whenever debug mode is enabled, which can disclose generated audio prompt text, model outputs, error details, or echoed sensitive inputs. Given this skill’s personalization features and handling of potentially sensitive user material, such logging increases the chance of privacy leakage through local logs and debugging output.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The document explicitly directs the skill to collect personalized context and then use it to compose an API-bound prompt, but it does not require a clear, up-front user-facing notice at the point of collection explaining what categories of personal data may be accessed and used. Even if processing is local, undisclosed collection and transformation of personal context can violate user expectations and lead to sensitive details influencing third-party output.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The prescribed voice options are all Chinese/Mandarin-specific, and later sections require Chinese-specific prompt fields such as Chinese labels and character limits. This creates a language/locale constraint without explicit user opt-in or a clearly documented justification for being region-specific.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

This section instructs the system to use specific details, emotional state, names, projects, conversation history, and no-go zones to personalize output, which materially increases sensitivity because it builds derived inferences about the user. The guidance to use 'fuzzy resonance' reduces overt disclosure but does not mitigate the underlying privacy risk of harvesting intimate contextual signals without explicit, granular permission.

Content

No source excerpt is available for this finding.