Back to skill

Security audit

Memory Read

Security checks for vulnerabilities and agentic risk

Overview

This skill is a disclosed memory-reading guide, but it allows access to private reports and archived chats with vague conversational triggers and weak scoping.

Review this before installing if the agent may access private work logs or archived chats. It should be used only in an environment where the available memory directories are already limited to data the current user is allowed to see, and cold chat retrieval should require explicit scope and consent.

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:17
Finding

Insufficient Authorization Controls for Sensitive Memory and Archived Chat Access

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 17–22 and 49–56
Vulnerability Type: Sensitive local-data access without enforceable identity or authorization validation
Risk Level: Medium

Relevant Source Snippet (English translation):

markdown
| Trigger scenario | Data to read | Purpose |
|------------------|--------------|---------|
| Rebuild after memory loss | Daily reports from the last seven days | Rebuild work context |
| Start of each day | Today's daily report | Determine what work should continue |
| Locate historical information | Index for the specified date | Locate the target file |
| The named user asks about progress | Corresponding personal directory | Check the named user's progress |
| Retrieve shared knowledge | Hot/warm storage | Obtain reference content |
| Retrieve chat history | Cold storage | Triggered only by an instruction from the named user, manually, no more than 50 KB |

## Cold-Storage Read Rules

- Write without automatic retrieval: Do not proactively read raw chats from cold storage.
- Trigger: The named user says “retrieve...”
- Safety limit: No more than 50 KB per operation; read in segments.
- After vector search becomes available: Enable full-text retrieval as a future capability.

Technical Analysis

The Skill permits access to daily reports, personal progress directories, indexed historical files, and raw archived conversations. Its authorization model relies on conversational recognition of a particular user name or trigger phrase rather than an authenticated identity, explicit owner consent, or an enforceable access-control decision.

The documented 50 KB read limit is a data-volume restriction rather than an authorization control. It does not prevent an unauthorized requester from initiating the read, and the documentation does not impose a cumulative limit that would prevent repeated segmented requests. The index-based lookup process also ...[truncated 1687 chars]

Remediation
View remediation

Remediation Suggestions

  1. Require authenticated caller identity and an explicit authorization decision before reading personal directories, historical memory, or archived chats. Do not treat a claimed name or trigger phrase as proof of identity.
  2. Require explicit, scope-specific owner consent for raw conversation retrieval, including the requested date range, data category, and intended recipient.
  3. Define an allowlist of permitted directories and canonicalize each resolved path before access. Reject absolute paths, traversal components, symlink escapes, and index entries that resolve outside approved roots.
  4. Apply least-privilege access controls so the Skill can read only the minimum records needed for the current authorized task.
  5. Add content filtering and redaction for credentials, secrets, personal data, and information belonging to unrelated third parties before returning retrieved content.
  6. Replace the per-request-only limit with cumulative limits per user, session, and time window. Require renewed authorization when the approved scope is exceeded.
  7. Record tamper-resistant audit events containing the verified requester, authorization result, data category, approved scope, files accessed, and amount returned.
  8. Fail closed when identity, consent, ownership, or path scope cannot be verified.
Vulnerability Patterns
  • 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
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (3)

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The activation conditions are described as broad situations such as memory reconstruction, daily start, or asking for progress, rather than precise, verifiable conditions. In a skill that governs memory reads, this ambiguity can lead to unnecessary loading of historical content, privacy overreach, or misuse by downstream agents interpreting normal conversation as authorization to access memory.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The trigger for cold-area/chat retrieval is underspecified as a simple phrase like '调取...,' which can cause accidental or unauthorized retrieval of sensitive archived conversations. In a memory-access skill, vague invocation criteria increase the chance that an agent will over-read private data based on ambiguous user wording or prompt manipulation.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Low
Category
Not specified by scanner
Confidence
84% confidence
Finding

Line L76 states the module is 'not active loading', which suggests it does not perform loading behavior itself. However, the rest of the document defines concrete triggers and procedures for reading memory, including manual retrieval of cold-zone chat data at L24 and L58-L63, so the statement understates and partially contradicts the module's actual documented role.

Content

No source excerpt is available for this finding.