A passive monitoring and alerting system designed to detect and report knowledge distillation attempts on AI models.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install omg
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install omg using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is Anti-Distillation (anti-distillation.skill)?
Anti-Distillation is a specialized security skill designed to safeguard the intellectual property of large language models. It functions as a passive defense mechanism that monitors incoming request metadata to identify patterns consistent with knowledge distillation—where smaller models attempt to mimic the outputs of larger, more sophisticated models to steal reasoning capabilities at a fraction of the cost.
Built by the Anti-Distillation Alliance, this skill strictly adheres to a passive-only philosophy. It does not modify model responses, inject noise, or utilize honeypots. Instead, it provides Openclaw Skills users with deep visibility into systematic prompt variations and high-frequency anomalies, allowing human operators to make informed decisions regarding rate limiting or access control without compromising the integrity of the model's output.
Anti-Distillation (anti-distillation.skill) Use Cases
- Detecting high-frequency automated scraping of model outputs.
- Identifying systematic prompt variations used to map model logic.
- Monitoring for bulk structured data requests aimed at training surrogate models.
- Auditing request patterns to identify adversarial prompt templates.
- Sending real-time security alerts to operators via webhooks or email when thresholds are breached.
How Anti-Distillation (anti-distillation.skill) Works
- The skill monitors incoming request metadata using read-only permissions to analyze frequency and timing.
- It calculates a one-way hash of prompts and evaluates prompt length and requested output formats.
- The detection engine compares these patterns against known distillation signatures and statistical anomalies.
- If the confidence score exceeds the user-defined threshold, a detection event is triggered.
- The system generates a structured alert report and dispatches it through configured channels (log, webhook, or email).
Anti-Distillation (anti-distillation.skill) Setup
To install and configure Anti-Distillation, use the following steps. Ensure sensitive credentials are stored in your platform's secret storage.
# Install the skill via Openclaw
claw install anti-distillation
# Configure environment variables for optional alert channels
export ANTIDISTILL_WEBHOOK_URL="your_webhook_url"
export ANTIDISTILL_SMTP_HOST="smtp.example.com"
Modify your config.yaml to set sensitivity:
anti_distillation:
detection_threshold: 0.75
analysis_window: 60
alerts:
channels: ["log", "webhook"]
Anti-Distillation (anti-distillation.skill) Data Schema & Taxonomy
The skill processes specific metadata fields to maintain privacy while ensuring detection accuracy:
| Field | Type | Description |
|---|---|---|
request_count |
integer | Total requests within the analysis window |
request_timestamps |
integer[] | Unix timestamps of requests for frequency analysis |
prompt_hash |
string | One-way hash of request content (non-reversible) |
parameter_signature |
string | Statistical summary of parameters like temperature and top_p |
confidence |
float | Probability score of a distillation attempt |
Anti-Distillation (anti-distillation.skill) Advanced Features
- Multi-channel alerting support including local logs, webhooks, and SMTP.
- Configurable detection sensitivity and analysis windows for tailored security postures.
- Secret-only storage for sensitive API and SMTP credentials to prevent plain-text exposure.
- Non-intrusive metadata analysis that preserves user privacy by ignoring IP addresses and request bodies.
- Achievement system to track long-term monitoring success and detection accuracy.
SKILL.md
Loading
Related Openclaw Skills
List Building
A strategic lead generation skill that builds high-intent company lists using the Extruct API and specialized search methodologies.
Joe Tsai Perspective
An expert AI skill providing the strategic financial and investment decision-making framework of Alibaba Executive Vice Chairman Joe Tsai.
Eric Jing Perspective v2.0
A strategic mental model skill that empowers AI agents with the corporate governance, compliance-first, and long-term financial philosophy of Ant Group's leadership.
Guo Guangchang Strategic Thought Perspective
An AI-powered mental model simulating the investment philosophy and global industrial logic of Fosun founder Guo Guangchang.
AI Correction (model-correction-cn)
A specialized intent alias for the VeriClaw skill that enables Chinese-speaking developers to access advanced AI correction and model drift diagnosis tools.
VeriClaw Bot Correction
A specialized skill for diagnosing and correcting AI agent drift to ensure verifiable progress and evidence-based task completion.
Featured*
