Back to skill

Security audit

check-notes-first

Security checks for vulnerabilities and agentic risk

Overview

The skill is coherent and not malicious, but it directs agents to search broad memory and conversation history without clear user control or sensitivity limits.

Install only if you are comfortable with the agent consulting prior memory, notes, scripts, wikis, and possibly conversation history for relevant methods. Prefer using it in environments where those stores are scoped to you and do not contain credentials or unrelated private material.

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:47
Finding

Overbroad Access to Agent Memory and Conversation History

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 47–71
Vulnerability Type: Excessive access to potentially sensitive historical data
Risk Level: Medium

Vulnerable Code

markdown
1. curated operational notes and durable memory;
2. recent dated logs or notes;
3. reusable scripts/configuration;
4. searchable conversation/session history;
5. compiled wiki or semantic/vector recall.

Search by **operation, mechanism, and synonyms**.

Examples:

- `mmproj` → `vision projector`, `multimodal`, `attach vision`, `files map`;
- `systemd timer broke` → `unit`, `schedule not firing`, `cron`, `guard script`;
- `patch did not survive upgrade` → `replay`, `renamed chunk`, `reinstall clobbered`.

Try up to three materially different queries by default. Continue only if a result exposes a
concrete new lead, or the risk justifies deeper retrieval. Otherwise use first principles.

### Known fast paths when present

- `TOOLS.md` — durable operational how-tos and machine specifics;
- `MEMORY.md` — curated long-term context and guardrails;
- `memory/YYYY-MM-DD.md` — dated raw logs and recent methods;
- `memory_search` with `corpus=all` — durable + semantic/vector recall;
- `scripts/` — reusable solutions; search for the operation, not only the product name;
- session/conversation logs — where the original solve may have happened;
- wiki/search tools — accumulated syntheses.

Technical Analysis

The Skill directs the agent to search broad historical data sources, including durable memory, dated logs, memory_search with corpus=all, and complete session or conversation history. It does not require task-specific authorization, source allowlisting, sensitivity classification, redaction, or user consent before accessing these stores.

Such repositories can contain credentials, personal information, confidential operational details, or content belonging to unrelated users or sessions. Searching ...[truncated 1894 chars]

Remediation
View remediation

Remediation Suggestions

  1. Limit retrieval to task-scoped and explicitly approved knowledge sources rather than searching all available memory.
  2. Remove corpus=all as a default fast path. Require a narrowly selected corpus and task-specific query.
  3. Exclude session and conversation logs by default. Require explicit user consent or a documented authorization basis before accessing them.
  4. Define path and source allowlists for operational notes, scripts, and memory files. Deny access to credential stores, private conversations, and unrelated user data.
  5. Apply secret and personal-data detection before retrieved content is inserted into the active context, displayed in output, or persisted.
  6. Retrieve only the minimum relevant excerpt and avoid copying raw records when a sanitized method summary is sufficient.
  7. Add a rule prohibiting credentials, tokens, personal data, and unrelated confidential details from being written into reusable notes.
  8. Preserve provenance and access-control metadata so the agent can reject records that are not authorized for the current user or task.
  9. Audit memory searches and write-back operations, including the selected source, query scope, reason for access, and redaction outcome.
  10. Prefer first-principles analysis whenever authorized, sanitized prior knowledge cannot be located safely.
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep