Back to skill

Security audit

Personal Os Framework

Security checks for vulnerabilities and agentic risk

Overview

This documentation-only skill is coherent, but it asks an AI to continuously read and update a personal knowledge base without clear approval, privacy, or rollback safeguards.

Install only if you want an AI to maintain a dedicated personal-os workspace. Keep it confined to that workspace, use version control or backups, avoid storing secrets or highly sensitive personal data, require diffs or confirmation before writes, and do not enable cron-style maintenance with broad filesystem access.

Vulnerability Patterns
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • 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 (1)

T05 · Unauthorized Access and Privilege Escalation

Warning
Location
SKILL.md:70
Finding

Unbounded Automatic Persistence of Conversation-Derived Data

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:70-87; related instructions in docs/setup.md:13-23 and docs/setup.md:29-41
Vulnerability Type: Standing filesystem read/write authority without explicit consent or scope controls
Risk Level: Medium

Vulnerable Code Snippets

SKILL.md:70-87:

markdown
The AI does not wait for commands. The AI reads the system, understands the human, and acts proactively.

Reviews keep the system alive. Without regular reviews, the system becomes stale.

Decisions are the most valuable content. Log them with context.

## Files

Every personal-os has these core files:
- `STATE.md` — current project states
- `TODO.md` — active tasks
- `DECISIONS.md` — decision log
- `PENDING-DECISIONS.md` — blocked decisions
- `HEARTBEAT-LOG.md` — system activity log

## For AI Collaborators

Read the personal-os after each conversation. Update relevant files. The system stays current because the AI keeps it current.

docs/setup.md:13-23:

markdown
3. **Start working with your AI collaborator**
4. **Update the system after each significant conversation or action**

## For the AI

After each conversation:

1. Read `STATE.md`, `DECISIONS.md`, `TODO.md`
2. Update relevant files with new decisions, progress, or changes
3. Write a summary to `HEARTBEAT-LOG.md`
4. Generate review drafts periodically

docs/setup.md:29-41:

markdown
## Cron Jobs (Optional)

Set up periodic health checks and review generation:

- Every 30 min: consistency check
- Every 1 hour: system maintenance
- Every day: review generation
- Every week: deeper review

## No Command Required

The AI should not wait for commands.

The AI reads the system, understands the human, and acts proactively.

Technical Analysis

The Skill gives the agent standing authority to read persistent project records and write conversation-derived information after every ...[truncated 3039 chars]

Remediation
View remediation

Remediation Suggestions

  1. Require explicit consent for persistence

    • Ask for confirmation before writing conversation-derived information.
    • Present the target file and exact proposed diff before applying changes.
    • Permit automatic writes only after the user explicitly enables them for a defined scope.
  2. Enforce filesystem boundaries

    • Configure one canonical workspace root.
    • Resolve and validate paths before every operation.
    • Reject absolute paths, parent-directory traversal, and symbolic-link escapes.
    • Use an allowlist for files and directories the Skill may read or modify.
  3. Apply data-minimization rules

    • Do not persist passwords, API keys, authentication tokens, financial information, health data, or other sensitive personal information.
    • Redact likely secrets before proposing a write.
    • Store only the minimum context necessary for the requested task.
  4. Separate trusted instructions from captured content

    • Treat imported notes, webpages, emails, and document contents as untrusted data.
    • Never convert instructions found inside captured content into agent policies or autonomous actions.
    • Require user confirmation before promoting untrusted material into durable decision or task records.
  5. Add integrity and recovery controls

    • Use atomic writes and maintain version history.
    • Record the source, timestamp, and approving user for each update.
    • Provide simple rollback and deletion procedures.
    • Avoid overwriting existing entries when an append-only record is sufficient.
  6. Constrain unattended maintenance

    • Disable periodic maintenance by default.
    • Require explicit configuration for every scheduled operation.
    • Run maintenance with least-privilege access limited to the approved workspace.
    • Generate review drafts without automatically changing authoritative project state.
  7. **Clarify the Skill instruct ...[truncated 284 chars]

Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Memory PoisoningPersistent Context Injection, Context Window Stuffing, Memory Manipulation
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (15)

Missing User Warnings

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The instruction that the AI should not wait for commands and should act proactively encourages autonomous behavior without any stated authorization boundary, approval workflow, or limitation on what actions are permitted. Because this skill's purpose is to read and maintain a personal operating system, the context makes the issue more dangerous: proactive writes across decision logs, task files, and reviews can silently reshape trusted personal records and trigger cascaded automation from inaccurate updates.

Content

No source excerpt is available for this finding.

Memory Manipulation

High
Category
Memory Poisoning
Confidence
80% confidence
Finding

Skill manipulates agent memory, state, or stored context. Memory corruption can alter personality, override safety rules, or cause unpredictable behavior.

Content

Scanner excerpt · docs/task-layer-guide.md (reported line 5)May include surrounding context.

md
## What is a Task?

A task is a unit of work with a clear owner and a clear state.

Not everything is a task. Meeting notes, decisions, reference material — these are not tasks.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The README encourages an AI to read, maintain, and act proactively on a user's personal knowledge base without any privacy, consent, retention, or access-control warnings. In a framework explicitly designed to store decisions, tasks, and contextual personal data, this omission can lead users to expose sensitive information to an AI system or connected tooling without understanding the risks.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The setup instructions tell users to 'let your AI collaborator record your decisions' and maintain core files, but provide no warning about data integrity, prompt manipulation, or accidental modification of important records. Because this skill's purpose is ongoing AI collaboration over user-maintained state, missing safeguards make unintended disclosure, corruption, or over-collection of personal information more likely.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The skill explicitly instructs the AI to read and update user-maintained files after each conversation, but it provides no guardrails around consent, scope limitation, sensitive data handling, or confirmation before modifying personal records. In a personal knowledge-management context, this can lead to unintended overwrites, privacy exposure, or silent corruption of a user's notes and task system, especially because the framework encourages proactive behavior rather than explicit user-triggered actions.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The document explicitly states that the AI 'does not wait for commands' and should 'act proactively,' but it provides no guardrails about confirmation, scope limits, or the kinds of actions the AI may take on user data or connected systems. In a framework designed for ongoing reading and updating of a personal operating system, this can normalize autonomous behavior that causes unintended state changes, privacy exposure, or unsafe actions without clear user consent.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The overview repeatedly frames the AI as acting proactively rather than waiting for commands, culminating in instructions to continuously read and maintain the system. Without clear constraints on when the skill activates, what actions are allowed, and how user consent is obtained, this encourages autonomous behavior that can exceed user intent and cause unauthorized updates or decisions based on incomplete context.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The document instructs the AI to read the personal OS after each conversation and update relevant files, but it does not require explicit user confirmation, define write boundaries, or warn that project data will be modified. In a system described as a persistent 'second brain,' this can lead to unintended or overbroad state changes, corruption of notes/tasks, or silent persistence of sensitive information.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

This markdown file contains copy-paste prompts that direct the AI to create project structure and update STATE.md and TODO items, which are user-data modifications. The document does not provide any warning or disclosure that using these prompts may create or alter files in the user's personal-os workspace.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The prompt instructs the AI to write a weekly review directly to Chronicle/reviews/[date]-weekly.md without warning that this will create or overwrite a local file. In an agentic environment, that can lead to unintended persistence, overwriting prior notes, or silent data modification if the user treats the prompt as read-only analysis.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

This prompt directs the AI to classify captured notes and save them into multiple files or folders without telling the user that their content will be moved, copied, or stored persistently. That increases the risk of accidental disclosure, misrouting sensitive information, or unwanted retention of notes in durable locations.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The monthly reflection prompt instructs the AI to write output to Chronicle/reviews/[date]-monthly.md without disclosing file creation or modification behavior. This can surprise users, create unintended records, or overwrite existing reflections when run repeatedly in automated settings.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The quickstart explicitly instructs the AI to write to and update local user files such as DECISIONS.md, STATE.md, and TODO.md without any safety guidance, confirmation step, or boundary on what files may be modified. In an agentic setting, normalizing autonomous file writes can lead to unintended data modification, overwrite of important notes, or broader misuse if the agent has filesystem access and the user assumes examples are safe defaults.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

Telling the user to update the system after each significant conversation or action normalizes continuous autonomous state changes without warning about the persistence and frequency of file modifications. In practice, this can lead to excessive or mistaken writes, propagation of hallucinated information into durable records, and user surprise about how much data is being changed.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The guide explicitly instructs the AI to read and then modify several user files after each conversation, but it provides no guardrails such as confirmation prompts, scope limits, backups, or change visibility. In a skill designed for ongoing personal knowledge management, this creates a real risk of unintended overwrites, silent data corruption, or writes to sensitive files if the workspace boundaries are unclear.

Content

No source excerpt is available for this finding.