Back to skill

Security audit

self-improving

Security checks for vulnerabilities and agentic risk

Overview

This skill is mostly a self-improvement memory system, but it asks for broad persistent memory and agent-control authority that users should review carefully.

Install only if you intentionally want a persistent self-improvement memory system. Review and limit what it can write, avoid global hooks, require confirmation before exporting, deleting, sharing across sessions, creating cron jobs, or modifying AGENTS.md/SOUL.md/TOOLS.md, and periodically audit stored memory for sensitive or unwanted content.

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
  • System PersistenceInstalls backdoors, hooks, services, or scheduled tasks that survive the run
  • 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)

T01 · Skill Instruction Hijacking

Error
Location
scripts/extractor.py:125
Finding

Untrusted Markdown content can be converted into authoritative agent instructions

Content
View full analysis
str: if not rules: return "" lines = ["Historical experience rules — follow these rules"] avoid_rules = [r for r in rules if r.category == "avoid"] prefer_rules = [r for r in rules if r.category == "prefer"] if avoid_rules: lines.append("\n## Operations to avoid") for r in sorted(avoid_rules, key=lambda x: -x.confidence): lines.append(f"- {r.description} (confidence: {r.confidence:.0%})") if prefer_rules: lines.append("\n## Recommended practices") for r in sorted(prefer_rules, key=lambda x: -x.confidence): lines.append(f"- {r.description} (confidence: {r.confidence:.0%})") return "\n".join(lines) ``` The labels in the second excerpt are English translations of the original user-facing strings; the executable interpolation of `r.description` is unchanged. ### Technical Analysis `extract_from_learnings()` accepts a caller-provided directory, reads every Markdown file in it, and ex ...[truncated 2361 chars]
Remediation
View remediation

T02 · Agent Memory Poisoning

Error
Location
SKILL.md:654
Finding

Learned content can be promoted into persistent agent-control files

Content
View full analysis
Remediation
View remediation

T06 · System Persistence

Warning
Location
SKILL.md:1359
Finding

Optional behavior is made persistent through scheduled tasks and agent bootstrap steering

Content
View full analysis
Remediation
View remediation
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • System Prompt LeakageDirect Leakage, Indirect Extraction, Tool-Based Exfiltration
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
Findings (38)

Intent-Code Divergence

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The skill’s stated scope says it only reads its own memory files, but other sections authorize reading broader session history and additional files. This mismatch is dangerous because operators may trust the narrow scope declaration while the actual instructions expand data access beyond what was disclosed, increasing privacy and data-exposure risk.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

High
Category
Not specified by scanner
Confidence
96% confidence
Finding

The document claims learning occurs only from explicit corrections, but later sections instruct capture of broad message content and even reinforcement from lack of correction. That creates undisclosed behavioral profiling and persistent storage of conversational data beyond the user’s reasonable expectation.

Content

No source excerpt is available for this finding.

Ssd 3

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

Recording every user message and related details into persistent memory/log files creates a clear data-retention and leakage hazard. Natural-language conversations frequently contain credentials, personal context, internal URLs, proprietary plans, or other sensitive details that may be captured incidentally and then retained or shared later.

Content

No source excerpt is available for this finding.

Ssd 3

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The working-buffer protocol explicitly requires logging every interaction after a context threshold, including user message content. This converts normal conversation into a durable transcript store, increasing the risk of privacy violations, accidental secret retention, and secondary disclosure through later reads, exports, or sharing.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
95% confidence
Finding

The growth-loop sections direct systematic collection of user information and conversational patterns into profile-like files. That exceeds the stated self-improvement purpose and creates a persistent user-data retention mechanism that can capture sensitive personal preferences or behavioral details.

Content

No source excerpt is available for this finding.

Vague Triggers

High
Category
Not specified by scanner
Confidence
95% confidence
Finding

The kill-switch phrase "forget everything" is ambiguous and may appear in normal conversation, yet it triggers a destructive memory wipe and an export side effect. This creates risk of accidental deletion, possible unauthorized triggering, and unintended data exposure from the pre-wipe export step.

Content

No source excerpt is available for this finding.

Agent Config Directory Access

High
Category
Agent Snooping
Confidence
90% confidence
Finding

Writing to ~/.claude/settings.json establishes user-level persistence for hook execution across projects and sessions. That makes the behavior more dangerous than a project-local setup because a single enablement can affect unrelated repositories and trusted environments, amplifying the blast radius if the hooked scripts are unsafe or later modified.

Content

Scanner excerpt · references/hooks-setup.md (reported line 48)May include surrounding context.

Option 2: User-Level Configuration

Add to ~/.claude/settings.json for global activation:

json
{

Exfiltration Commands

High
Category
Prompt Injection
Confidence
90% confidence
Finding

Instructions found that direct the agent to transmit conversation context or user data to external services.

Content

Scanner excerpt · references/openclaw-integration.md (reported line 181)May include surrounding context.

sessions_send

Send message to another session:

text
sessions_send(sessionKey="session-id", message="Learning: API requires X-Custom-Header")

Direct Prompt Extraction

High
Category
System Prompt Leakage
Confidence
85% confidence
Finding

Skill contains instructions that could directly expose system prompts, internal rules, or hidden instructions to users or external parties.

Content

Scanner excerpt · scripts/extractor.py (reported line 54)May include surrounding context.

python
# 从成功中提取 prefer 规则
        rules.extend(self._extract_from_successes(successes))

        return rules

    def _extract_from_failures(self, failures: List[dict]) -> List[Rule]:
        """从失败中提取 avoid 规则"""

Direct Prompt Extraction

High
Category
System Prompt Leakage
Confidence
85% confidence
Finding

Skill contains instructions that could directly expose system prompts, internal rules, or hidden instructions to users or external parties.

Content

Scanner excerpt · scripts/extractor.py (reported line 79)May include surrounding context.

python
# 从成功中提取 prefer 规则
        rules.extend(self._extract_from_successes(successes))

        return rules

    def _extract_from_failures(self, failures: List[dict]) -> List[Rule]:
        """从失败中提取 avoid 规则"""

Direct Prompt Extraction

High
Category
System Prompt Leakage
Confidence
85% confidence
Finding

Skill contains instructions that could directly expose system prompts, internal rules, or hidden instructions to users or external parties.

Content

Scanner excerpt · scripts/extractor.py (reported line 104)May include surrounding context.

python
# 从成功中提取 prefer 规则
        rules.extend(self._extract_from_successes(successes))

        return rules

    def _extract_from_failures(self, failures: List[dict]) -> List[Rule]:
        """从失败中提取 avoid 规则"""

Direct Prompt Extraction

High
Category
System Prompt Leakage
Confidence
85% confidence
Finding

Skill contains instructions that could directly expose system prompts, internal rules, or hidden instructions to users or external parties.

Content

Scanner excerpt · scripts/extractor.py (reported line 118)May include surrounding context.

python
# 从成功中提取 prefer 规则
        rules.extend(self._extract_from_successes(successes))

        return rules

    def _extract_from_failures(self, failures: List[dict]) -> List[Rule]:
        """从失败中提取 avoid 规则"""

Direct Prompt Extraction

High
Category
System Prompt Leakage
Confidence
85% confidence
Finding

Skill contains instructions that could directly expose system prompts, internal rules, or hidden instructions to users or external parties.

Content

Scanner excerpt · scripts/extractor.py (reported line 146)May include surrounding context.

python
# 从成功中提取 prefer 规则
        rules.extend(self._extract_from_successes(successes))

        return rules

    def _extract_from_failures(self, failures: List[dict]) -> List[Rule]:
        """从失败中提取 avoid 规则"""

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The skill is presented as self-reflection and memory logging, but it expands into unrelated operational capabilities like fleet analysis, deployment gating, cron management, and proactive growth tracking. This overbreadth increases attack surface and makes it harder for users to understand what the skill may do or authorize.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The opening description does not clearly warn that the skill persistently stores user-related information and conversation-derived data. This undermines informed consent, especially because later sections introduce extensive logging and memory retention behaviors.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

Broad trigger conditions can cause the skill to activate during ordinary conversations, leading to unexpected logging and memory updates. In a persistence-oriented skill, accidental activation is more dangerous because it can silently retain user-derived data over time.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

Cross-session transcript access and message sending introduce lateral data movement not strictly necessary for self-improving memory. Even with sanitization guidance, this creates a path for unintended disclosure of context across sessions and broadens the blast radius of any sensitive information captured earlier.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

Ambiguous natural-language invocation phrases overlap with normal user speech, increasing the chance of accidental execution of memory and export actions. Because this skill stores and bundles persistent data, false activations could expose or archive more user information than intended.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The export-memory function bundles all stored memory into an archive without a prominent privacy warning. Users may trigger export expecting convenience, but the archive could include sensitive conversation-derived notes, preferences, and logs accumulated over time.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
87% confidence
Finding

Later sections add more generic triggers without clear boundaries, making the skill’s activation model difficult to predict. Unclear triggers are especially risky here because the resulting actions include durable logging, self-modification workflows, and broader context collection.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

These sections instruct pervasive logging of user messages and responses without a clear privacy warning proportionate to the retention behavior. Persistent capture of ordinary interaction content creates a significant risk of storing secrets, personal data, or sensitive contextual details despite generic admonitions not to log secrets.

Content

No source excerpt is available for this finding.

Indirect Prompt Extraction

Medium
Category
System Prompt Leakage
Confidence
85% confidence
Finding

Skill contains patterns that could indirectly extract system prompts through rephrasing, translation, summarization, or side-channel techniques.

Content

Scanner excerpt · SKILL.md (reported line 1352)May include surrounding context.

md
---

## Part 30: Reverse Prompting(合并自 proactive-agent v3.1.0)

> 主动询问用户需要什么,而不是等待被告知。

Ssd 3

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The growth-loop instructions encourage systematic profiling of the user and storage of those learnings in a dedicated file. Even if intended to improve assistance, this creates a persistent behavioral dossier that is not necessary for the core skill and may capture sensitive personal preferences or patterns.

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 · boundaries.md (reported line 11)May include surrounding context.

md
| Financial | Card numbers, bank accounts, crypto seeds | Fraud risk |
| Medical | Diagnoses, medications, conditions | Privacy, HIPAA |
| Biometric | Voice patterns, behavioral fingerprints | Identity theft |
| Third parties | Info about other people | No consent obtained |
| Location patterns | Home/work addresses, routines | Physical safety |
| Access patterns | What systems user has access to | Privilege escalation |

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The phrase triggering a full memory export is very broad and could be invoked during ordinary discussion rather than as a deliberate privacy request. Because the action exposes all stored user data, accidental activation could cause unintended disclosure of sensitive memory contents.

Content

No source excerpt is available for this finding.