DarKnight
A self-improving agent framework for dark web threat intelligence scenarios. It employs dual-signal RAG retrieval, an image-specific preprocessing pipeline, and site feedback learning, running open-weight VLMs on local GPUs to solve CAPTCHA challenges.
Timeline (last 90 days)
DarKnight is a self-improving agent framework designed to run a locally hosted open-weight VLM on a single workstation GPU within investigators' own Tor crawling environment, without relying on third-party services and without requiring separate training for each market.
DarKnight employs a dual-signal RAG retrieval mechanism whose retrieval targets include both historical success cases and failure cases.
DarKnight dynamically combines targeted preprocessing schemes based on the characteristics of the current image, rather than applying a uniform pipeline to all images.
In case studies on four real dark web markets, the core conclusion is that the agent's tool selection capability is more effective than simply scaling up model parameters.
On one of the most heavily defended dark web markets, a deployable open-weight VLM combined with DarKnight achieved an 83.2% pass rate, 37.2 percentage points higher than its base model.
DarKnight's 83.2% pass rate with an open-weight VLM exceeds Claude Opus 5's 44% on the same task.
At a performance degradation cost of 7.6%, DarKnight recovered 48.2% of baseline failure cases.
On specific vertical tasks, a retrieval-guided agent framework paired with a small model may be more practical and sustainable than directly calling a large model API.
All Facts (8)
DarKnight is a self-improving agent framework designed to run a locally hosted open-weight VLM on a single workstation GPU within investigators' own Tor crawling environment, without relying on third-party services and without requiring separate training for each market.
50%UnverifiedDarKnight employs a dual-signal RAG retrieval mechanism whose retrieval targets include both historical success cases and failure cases.
50%UnverifiedDarKnight dynamically combines targeted preprocessing schemes based on the characteristics of the current image, rather than applying a uniform pipeline to all images.
50%UnverifiedIn case studies on four real dark web markets, the core conclusion is that the agent's tool selection capability is more effective than simply scaling up model parameters.
50%UnverifiedOn one of the most heavily defended dark web markets, a deployable open-weight VLM combined with DarKnight achieved an 83.2% pass rate, 37.2 percentage points higher than its base model.
50%UnverifiedDarKnight's 83.2% pass rate with an open-weight VLM exceeds Claude Opus 5's 44% on the same task.
50%UnverifiedAt a performance degradation cost of 7.6%, DarKnight recovered 48.2% of baseline failure cases.
50%UnverifiedOn specific vertical tasks, a retrieval-guided agent framework paired with a small model may be more practical and sustainable than directly calling a large model API.
50%