Back to skill

Security audit

Memory Landscape Review

Security checks for vulnerabilities and agentic risk

Overview

This skill is a straightforward memory-review helper that asks to read memory context and only proposes changes unless the user explicitly approves edits.

Before installing, understand that invoking this skill will cause the agent to inspect memory-related content, which may include private preferences or project instructions. Use it when you want that review, and ask the agent to list the memory locations first if you want tighter control over what is read.

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

Note
Location
SKILL.md:11
Finding

Insufficiently Scoped Access to Repository and Agent Memory

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:11, 26-27
Vulnerability Type: T05: Unauthorized Access and Privilege Escalation
Risk Level: Suspicious

Vulnerable Code

markdown
1. Gather repo memory files and current auto-memory context.
markdown
## Inputs
- Repo memory files
- Auto-memory state

Technical Analysis

The skill directs the agent to gather repository memory files and the current auto-memory state without defining path boundaries, permitted file types, exclusions, redaction requirements, or data-minimization controls. These memory layers may contain private preferences, prior-session information, sensitive operational instructions, or secrets accidentally recorded by users.

Access to memory is related to the skill's stated purpose, so the reviewed text does not demonstrate deliberate privilege escalation or malicious exfiltration. However, the unrestricted collection instruction may cause the agent to access and place more sensitive information into its active context or report than is necessary for a particular review.

The skill includes mitigating controls at SKILL.md:14-15 by requiring a grouped report and explicit approval before changes. Those controls restrict writes but do not adequately constrain the initial read and collection operation.

Attack Path

  1. A user invokes the memory landscape review skill.
  2. The agent interprets “repo memory files” and “current auto-memory context” broadly.
  3. The agent reads repository, local, personal, or auto-memory entries without first presenting an explicit access scope.
  4. Sensitive entries are loaded into the active model context.
  5. Those entries may be reproduced or summarized in the grouped report, exposing information beyond the user's intended review scope.

No instruction to transmit the collected information over a network, execute code, establish persistence, or alter memory without approval was foun ...[truncated 595 chars]

Remediation
View remediation

Remediation Suggestions

  1. Require the agent to enumerate candidate memory file paths before reading their contents.
  2. Obtain explicit user approval for the proposed read scope, especially for personal, local, and auto-memory.
  3. Define an allowlist of supported memory files and directories instead of using open-ended collection language.
  4. Exclude credentials, environment files, authentication tokens, private keys, and unrelated conversation records.
  5. Apply secret detection and redaction before memory content is included in model context or reports.
  6. Read only the minimum content required to classify each memory entry.
  7. Separate approval for repository-shared memory from approval for personal or auto-memory.
  8. Report sensitive entries by location and classification rather than reproducing their complete values.
  9. Retain the existing requirement that no promotions, cleanup, or other writes occur without explicit approval.
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