Back to skill

Security audit

Memory Tiers

Security checks for vulnerabilities and agentic risk

Overview

This memory-tier skill is review-worthy because it scans local agent transcripts, rewrites persistent memory files, and retains raw search queries in a shipped state file.

Install only if you are comfortable with this skill reading local OpenClaw session transcripts and modifying persistent memory files. Run maintenance in dry-run mode first, review proposed promotions/demotions carefully, delete or reset the bundled state/access-log.json before use, and avoid retaining raw search queries unless you explicitly want that telemetry.

Vulnerability Patterns
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • 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
  • 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 (2)

T02 · Agent Memory Poisoning

Warning
Location
scripts/track.js:117
Finding

Untrusted Transcript Content Can Trigger Persistent Memory Promotion

Content
View full analysis
promotionCutoff) { actions.push({ action: 'promote', section: section.title, from: tier, ...[truncated 3562 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
scripts/track.js:263
Finding

Plaintext Retention and Distribution of Sensitive Memory Search Queries

Content
View full analysis
100) { log.searches = log.searches.slice(-100); } } ``` Representative records distributed in `state/access-log.json:131-170`: ```json "searches": [ { "query": "agentview browser text-based token reduction", "timestamp": "2026-02-19T13:24:03.214Z" }, { "query": "job tracking list positions applications", "timestamp": "2026-02-20T22:24:41.364Z" }, { "query": "applications submitted last night February 19-20", "timestamp": "2026-02-20T15:01:00.925Z" }, { "query": "moltbook", "timestamp": "2026-02-21T00:40:38.862Z" } ] ``` ### Technical Analysis The access tracker persists complete memory-search query strings in plaintext, although the tier-management functionality only requires access timestamps and counters. Search queries can contain private topics, names, employment information, confidential project terms, credentials, or other sensitive data. The state file is stored inside the Skill project directory rather than a dedicated private runtime-data location. The audited artifact already includes populated search history, demonstrating that runtime behavioral data can be packaged or distributed with the Skill. Limiting retention to 100 entries reduces volume but does not address unnecessary collection, plaintext exposure, or artifact leakage. No network transmission or active exfiltration was found. Exposure occurs when the project directory or its state file is read, copied, backed up, comm ...[truncated 1323 chars]
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (15)

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

The description claims an automated tiered memory manager with LRU-style behavior and access-based promotion/demotion across hot/warm/cold tiers. The supplied code is only a manual command-line script for listing sections and moving them between three markdown files. It does not track access frequency/recency, does not automatically promote or demote items, and does not implement cache eviction or LRU logic. Its actual purpose is narrower and materially different: manual manipulation of markdown-backed memory tiers. Additionally, it performs filesystem reads/writes in the user's home workspace, which is an operational capability not reflected in the empty declared permissions.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
92% confidence
Finding

The declared description suggests an active tiered memory management system that implements LRU-like behavior across hot/warm/cold tiers, including automatic promotion, demotion, and access tracking. This code chunk only generates a 'Memory Tier Health Report' by reading existing memory tier files and an access-log JSON file, computing counts and staleness, and outputting recommendations. It does not track accesses itself beyond consuming prior tracking data, does not promote or demote content, does not enforce LRU behavior, and does not modify memory state. The code is therefore materially different in primary purpose and capabilities from the declared description.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The declared description suggests an active memory-management component that maintains hot/warm/cold tiers and automatically promotes or demotes items using LRU-style logic. The supplied code does not do that. Instead, it is an offline analytics/tracking utility: it scans session transcript JSONL files, extracts memory-related tool calls and tool outputs, records file/section access counts and timestamps, and saves an access log. While it does perform access tracking, that is only one small part of the declared functionality; the primary behavior is audit/log analysis rather than memory-tier management. The code also performs undeclared file-system scanning and persistence of tracking data, which are inconsistent with the stated purpose.

Content

No source excerpt is available for this finding.

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 41)May include surrounding context.

md
node scripts/maintain.js --dry-run # preview changes

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 42)May include surrounding context.

md
node scripts/maintain.js --dry-run # preview changes

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

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

md
node scripts/maintain.js --dry-run # preview changes

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 64)May include surrounding context.

md
node scripts/report.js # pretty-printed

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 65)May include surrounding context.

md
node scripts/report.js # pretty-printed

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
81% confidence
Finding

The skill instructs users to run local Node scripts that read and write memory and state files, but it does not declare any tool scope or permissions. That omission reduces transparency and weakens policy enforcement, making it easier for an agent or user to invoke filesystem-capable behavior without an explicit trust decision.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The skill explicitly recommends running maintenance commands that may modify memory files, but it does not prominently warn that content can be moved between tiers or changed on disk. In an agent-memory context, silent reorganization of workspace files can cause data loss, integrity issues, or unintended exposure of sensitive context to tiers with different loading/search behavior.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The script persistently mines agent session transcripts and stores derived telemetry about memory access, which goes beyond a narrowly described LRU-style memory tier manager. Even if intended for optimization, transcript mining and durable logging can capture sensitive usage patterns and create a secondary surveillance dataset that increases privacy and data-governance risk.

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 detailed access-tracking data to a local JSON file without any user-facing warning, consent flow, or clear retention policy. Silent persistence of behavioral metadata is dangerous because local logs are often overlooked, may have weak permissions, and can later be accessed by other tools or users on the system.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

This code parses broad session transcripts, extracts tool calls, and records search activity, allowing it to infer user interests and agent behavior far beyond cache management. Because transcripts may contain sensitive prompts, file paths, and contextual metadata, this creates unnecessary exposure and a privacy-sensitive audit trail that could be abused or leaked.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

Tracking writes to memory files and persisting file/section-level access telemetry creates a durable map of what information was touched and modified. In an agent environment, that metadata can reveal sensitive projects, priorities, or knowledge areas even without storing the full content, making it a meaningful privacy and operational-security concern.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

Persisting raw session search queries is especially sensitive because queries often directly encode user intent, secrets, investigative topics, or proprietary context. In a memory-tiering skill, this collection is poorly justified, and if the log is exposed it could disclose far more than simple cache statistics.

Content

No source excerpt is available for this finding.