Pith. sign in

REVIEW 1 cited by

WHITE PAPER: A Brief Exploration of Data Exfiltration using GCG Suffixes

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2408.00925 v1 pith:AHZ3JA3G submitted 2024-08-01 cs.CR cs.AI

classification cs.CRcs.AI
keywords dataattackexfiltrationsuffixxpiainjectionnearlyodds
verification ladder T0 review T1 audit T2 compute T3 formal
0 comments
abstract

The cross-prompt injection attack (XPIA) is an effective technique that can be used for data exfiltration, and that has seen increasing use. In this attack, the attacker injects a malicious instruction into third party data which an LLM is likely to consume when assisting a user, who is the victim. XPIA is often used as a means for data exfiltration, and the estimated cost of the average data breach for a business is nearly $4.5 million, which includes breaches such as compromised enterprise credentials. With the rise of gradient-based attacks such as the GCG suffix attack, the odds of an XPIA occurring which uses a GCG suffix are worryingly high. As part of my work in Microsoft's AI Red Team, I demonstrated a viable attack model using a GCG suffix paired with an injection in a simulated XPIA scenario. The results indicate that the presence of a GCG suffix can increase the odds of successful data exfiltration by nearly 20%, with some caveats.

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. The Resurgence of GCG Adversarial Attacks on Large Language Models

    cs.CL 2025-08 conditional novelty 3.0 of 10

    The paper reports that GCG attack success drops with model size, prefix-based metrics overestimate harm, and coding prompts are more vulnerable, but these findings rest on a small model set and a GPT-4o pipeline that ...

Pith tools