Smashing the token limit with overlapping fragments

* A researcher collaborated with Alex to exfiltrate larger tokens using tools like DOM Invader.
* The process involved stitching together smaller chunks to extract longer tokens without recursive loading.
* The method used graph theory to connect overlapping text fragments from a link into a complete token sequence.
* Tests demonstrated that using two-character and three-character chunks allowed for reconstruction, with certain grouping strategies yielding high success rates under specific thresholds (e.g., 90% or 99% probability).
* For a 12-character token, the shortest successful CSS required 1.33 MB, and extracting 640 characters within Gareth's original byte budget was achieved with up to five candidates.
* The final tested setup for achieving a 99% one-guess target for a 12-character token used 1.33 MB of CSS and 25,088 selectors.
* The technique was applied to generate test tokens by creating new URLs with randomized data for empirical validation.
* The process was demonstrated using Portswigger labs, where an AI tool was adapted to exploit CSS injection for token extraction across various web resources.

The core implication of this research is a shift from brute-force methods relying on sequential loading to an overlapping, combinatorial approach that leverages relational structure (graph theory) for data reconstruction. The initial breakthrough moves beyond simple exfiltration by treating the resulting code as a constrained system where the goal is not just extraction but efficient pathfinding through potential overlaps. The pattern of increasing chunk length and testing more complex joins reveals a tension between computational efficiency (minimizing CSS size) and epistemic certainty (ensuring the reconstructed token is correct). This exploration suggests that long-range data manipulation in web contexts may be less about raw brute force and more about exploiting the inherent, often overlooked, syntactic relationships within the injected code. The final results showing reductions in required CSS by orders of magnitude suggest a systemic vulnerability where complexity can be managed through structured modeling rather than sheer volume.
What assumptions about the efficiency of data extraction versus certainty are we making when prioritizing smaller stylesheets? How does this technique change the perceived security boundary between code injection and data exfiltration? Does focusing on constructing optimal join groups risk overlooking novel, more complex semantic relationships that might offer exponential gains in token recovery?