Back to skill

Security audit

dr-context-pipeline

Security checks for vulnerabilities and agentic risk

Overview

This memory helper needs review because casual continuation language can trigger lasting memory changes, while diagnostics can expose private context.

Require explicit save requests and review changes to standing rules. Redact diagnostics before displaying or retaining them, restrict access, and set a deletion period. Back up customized skill files before forced refreshes. Confirm which memory files and embedding provider your host uses; that configuration was not bundled. The supplied static scan reported no findings, but the inspected instructions support these review concerns.

Vulnerability Patterns
  • 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
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Vulnerability Patterns
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (10)

Skill declares no tool scope ('permissions' or 'allowed-tools') but code capabilities were detected: file_read, file_write, shell.

Medium
Category
MCP Least Privilege
Confidence
70% confidence
Finding

Without declared permissions the skill's intent is opaque and cannot be validated.

Content

No source excerpt is available for this finding.

Debug/audit artifact persistence lacks a warning about retaining private memory and user context.

Medium
Category
Output Handling
Confidence
89% confidence
Finding

Debug and audit modes require persisting retrieved memory and user context without specifying sanitization, access restrictions, or retention limits in this source. Because this pipeline handles persistent personal context, diagnostic artifacts can create additional copies of sensitive information exposed through workspace sharing, backups, or version control.

Content

No source excerpt is available for this finding.

Full diagnostic artifact output is required without warning that it can expose private memory.

Medium
Category
Output Handling
Confidence
93% confidence
Finding

The README requires full diagnostic artifacts to be persisted and emitted in debug or audit mode without a privacy filter. Requesting diagnostics does not necessarily authorize disclosing all retrieved private memory into a transcript that may have a broader audience or longer retention.

Content

No source excerpt is available for this finding.

Case-insensitive substring routing uses everyday words that can select unrelated retrieval workflows.

Medium
Category
Trigger Abuse
Confidence
97% confidence
Finding

The router uses first-match substring matching with broad terms such as "job", "missed", "tone", "plan", and "why". Ordinary messages can therefore select an unrelated task type; for example, "job interview" selects operations, while "planet" matches planning through "plan".

Content

No source excerpt is available for this finding.

“We’ll continue later” can trigger durable memory writes without an explicit request to save the session.

Medium
Category
Trigger Abuse
Confidence
95% confidence
Finding

“We’ll continue later” expresses an intention to resume, not necessarily consent to durable storage. Treating it as a save instruction can copy sensitive session content into multiple long-lived memory files and promote temporary discussion into standing rules.

Content

No source excerpt is available for this finding.

Debug and audit output exposes full retrieved context without a privacy warning.

Medium
Category
Output Handling
Confidence
95% confidence
Finding

Debug mode explicitly emits full Retrieval Bundle and Context Pack JSON, and audit mode inherits that disclosure. These objects can contain private memory unrelated to what the recipient needs, while the checklist specifies no redaction or audience-specific disclosure check.

Content

No source excerpt is available for this finding.

The installation workflow uses forced installation without warning that existing local skill files may be overwritten.

Medium
Category
Output Handling
Confidence
91% confidence
Finding

The prescribed install/refresh command uses --force without first checking for local modifications or preserving a backup. Applying this workflow can overwrite customized skill files, potentially removing local safeguards; displaying the workspace diff afterward does not prevent that loss.

Content

No source excerpt is available for this finding.

Everyday words and unrestricted substring matching can select unintended task workflows.

Medium
Category
Trigger Abuse
Confidence
98% confidence
Finding

The first-match router treats words such as job, missed, tone, and plan as sufficient triggers without domain or intent constraints. For example, a question about a job interview selects operations, and tone can match inside stone, selecting document review.

Content

No source excerpt is available for this finding.

The daily-log trigger includes a phrase that does not establish a request for historical recall.

Medium
Category
Trigger Abuse
Confidence
96% confidence
Finding

The comment limits daily-log activation to explicit time or recency requests, but did we does not reliably establish that intent. A conceptual question such as 'How did we get the term machine learning?' can therefore enable unrelated daily-log retrieval.

Content

No source excerpt is available for this finding.

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · scripts/memory_watchdog.py (reported line 87)May include surrounding context.

python
def run_json_command(command: list[str], timeout_seconds: int) -> tuple[Any | None, dict[str, Any]]:
    proc = subprocess.run(
        command,
        check=False,
        text=True,