Midos Self Improver

• Outcome: Review
• Latest audit: May 29, 2026 at 9:22 PM UTC
• Version:
  1.0.0

Overview

This skill is coherent with its self-improvement purpose, but it asks agents to persist learning data and promote rules into future agent instructions without enough user approval, redaction, or cleanup controls.

Install only if you want an agent to maintain project-local learning files and potentially change future agent instructions. Before using it, require manual approval for any promotion into CLAUDE.md, AGENTS.md, or permanent memory; review diffs; add secret redaction; and periodically prune the `.learnings`, `.patterns`, and `.knowledge` stores.

SkillSpector

By NVIDIA

Vulnerability Patterns

• Data Exfiltration: External Transmission, Env Variable Harvesting, File System Enumeration
• Trigger Abuse: Overly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
• Prompt Injection: Instruction Override, Hidden Instructions, Exfiltration Commands
• Privilege Escalation: Excessive Permissions, Sudo/Root Execution, Credential Access
• Supply Chain: Unpinned Dependencies, External Script Fetching, Obfuscated Code

Findings (3)

Vague Triggers

Medium

• Confidence: 84% confidence
• Finding: The correction detector uses very broad phrases like 'actually', 'that's wrong', and 'don't do that', which commonly appear in normal conversation and may trigger persistent learning capture when no true correction occurred. In this skill's context, false activations are more dangerous because they feed directly into long-lived memory and can eventually influence promoted rules, creating silent behavior drift or prompt-injection-style contamination of the knowledge base.

Vague Triggers

Medium

• Confidence: 87% confidence
• Finding: The knowledge-gap trigger fires when the agent says 'I don't know' or searches more than three times, but those conditions are vague and may reflect normal exploration rather than a genuine durable knowledge gap. Because the system stores and scores such events for later promotion, ambiguous triggering can accumulate noisy or misleading entries that distort future behavior and pollute project memory.

Missing User Warnings

Medium

• Confidence: 93% confidence
• Finding: The skill instructs persistent logging of corrections, tool commands, error messages, root causes, and fixes without any sanitization, minimization, or warnings about secrets and personal data. In practice, error output and user corrections often contain tokens, internal paths, stack traces, proprietary code details, or sensitive operational context, so this design can create a durable local data-exposure store that is later searched, promoted, or surfaced through connected tooling.

VirusTotal

64/64 vendors flagged this skill as clean.