Security audit

Hermesclawzero Auto Memory

Security checks for vulnerabilities and agentic risk

Overview

This memory skill is mostly purpose-aligned, but it asks for broad silent persistent memory, global reuse, background scheduled execution, and a Telegram reminder flow that is not clearly scoped or disclosed.

Review before installing. This skill will persistently store conversation-derived information, silently reuse it in later chats, and may create recurring jobs that run external local scripts. Only use it if you trust the sidecar scripts and database, understand what may be saved globally, and are comfortable with any Telegram reminder data flow; otherwise require explicit save/load notices, scoped memory, deletion controls, script integrity checks, and a documented way to disable the cron jobs.

A.I.G

Vulnerability Patterns

Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
System PersistenceInstalls backdoors, hooks, services, or scheduled tasks that survive the run
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

Error

Location: SKILL.md:25
Finding: Persistent Memory Poisoning Through Automatic Capture of User-Controlled Instructions
Content: View full analysis
"` - Optionally pass a `scope_id` (e.g. project name) for logical grouping. - Do **not** pass a `chat_id` — that makes memories unfindable via default search. Omitting it stores in `chat_id=global`. ``` ### Technical Analysis The skill instructs the agent to treat user-provided instructions as information eligible for immediate persistent capture. It then directs future sessions to retrieve and silently absorb stored context. The default storage behavior places captured information in the global `chat_id` scope. No controls are specified to: - Distinguish factual memories from behavioral instructions. - Mark stored content as untrusted data. - Confirm memory creation with the user. - Validate or sanitize content before persistence. - Isolate memories by conversation or project by default. - Prevent retrieved text from influencing agent instructions. - Display loaded memories so the user can identify poisoned state. - Expire, review, or delete unsafe entries. This creates a persistent trust-boundary failure. An attacker-controlled instruction can be summarized into long-term memory and later reintroduced into unrelated ses ...[truncated 1698 chars]
Remediation: View remediation

T06 · System Persistence

Error

Location: SKILL.md:45
Finding: Persistent Recurring Execution of Unreviewed External Scripts
Content: View full analysis
Remediation: View remediation

SkillSpector

Vulnerability Patterns

Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands

Findings (6)

Context-Inappropriate Capability

High

Category: Not specified by scanner
Confidence: 95% confidence
Finding: The Telegram push reminder capability expands the skill beyond local memory storage into outbound messaging, which materially changes the data exposure boundary. Even if the message content is limited, a reminder/push subsystem can leak memory-derived content, metadata, or behavioral patterns to external channels and is not justified by the stated purpose of local persistent memory. The context makes this more dangerous because the skill otherwise emphasizes local-only processing, creating a misleading trust boundary.
Content: No source excerpt is available for this finding.

Missing User Warnings

Medium

Category: Not specified by scanner
Confidence: 88% confidence
Finding: The skill instructs the agent to silently load stored user context at session start without a contemporaneous user-facing notice. This undermines informed consent and can cause the agent to act on historical personal data that the user does not expect is being retrieved in the current interaction. The risk is amplified in this context because the memory is persistent and intended to shape future behavior across sessions.
Content: No source excerpt is available for this finding.

Ssd 3

Medium

Category: Not specified by scanner
Confidence: 89% confidence
Finding: Silently ingesting stored user context without surfacing what was retrieved reduces transparency and prevents the user from correcting stale, sensitive, or incorrect memory before it influences agent behavior. This can lead to privacy harms, erroneous personalization, and inappropriate use of prior-session data. In a memory-management skill, hidden retrieval is particularly concerning because it normalizes unseen background profiling.
Content: No source excerpt is available for this finding.

Ssd 3

Medium

Category: Not specified by scanner
Confidence: 93% confidence
Finding: Automatically retaining and storing user-shared information across conversations in persistent memory is a genuine privacy and security risk when done by default. Cross-session persistence can accumulate sensitive profiles, instructions, and project details that become a high-value target and may be reused out of context. The skill context makes this more dangerous because the storage is continuous and broad rather than narrowly scoped to user-requested memory operations.
Content: No source excerpt is available for this finding.

Vague Triggers

Medium

Category: Not specified by scanner
Confidence: 95% confidence
Finding: The auto-save trigger is defined as capturing 'anything worth remembering,' including facts, instructions, preferences, and project details, which is too broad for safe autonomous storage. Ambiguous collection rules cause the agent to over-collect and persist sensitive user data without consistent necessity, increasing privacy risk and the likelihood of storing secrets or confidential business information. In a persistent memory skill, broad trigger language is especially risky because retention spans conversations.
Content: No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium

Category: Not specified by scanner
Confidence: 92% confidence
Finding: The skill’s privacy notice claims it captures only meaningful facts, but it also documents a full chat history backup flow. That creates a data minimization and transparency gap: users may reasonably expect summarized memory storage while the skill can persist substantially more sensitive conversational content on request. In a memory skill, this mismatch increases the chance of over-collection and retention of secrets, personal data, or regulated content.
Content: No source excerpt is available for this finding.

Static analysis

No suspicious patterns detected.