DarKnight: Retrieval-Guided Agentic Framework for Breaking Dark Web CAPTCHA Gates
DarKnight uses local open-weight vision models to solve dark-web CAPTCHAs for threat investigators.
DarKnight is a cs.CR paper describing a local agent that helps cyber-threat investigators pass dark-web marketplace CAPTCHAs without commercial vision-model APIs or per-site fine-tuning. It retrieves similar past successes and failures, builds an image-specific preprocessing pipeline, and uses the site’s pass or fail result. On four real marketplaces, an open-weight model with DarKnight reached 83.2% on the most defended site, 37.2 points above its base model and above Claude Opus 5 at 44%. It recovered 48.2% of baseline failures with a 7.6% degeneration cost.
- DarKnight runs small open-weight vision models locally inside a Tor crawler.
- On the hardest of four markets it reached 83.2%, 37.2 points above its base model.
- That score beat Claude Opus 5 at 44% without per-market training or commercial APIs.
- It recovered 48.2% of baseline failures at a 7.6% degeneration cost.
From the original article
1 min read at arxiv.orgCyber Threat Intelligence (CTI) investigators who monitor dark web marketplaces for leaked credentials and breach listings must pass the diverse CAPTCHAs that operators deploy to block automated collection.
By Binghui Wu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.