AuthRank: Scoring Source Credibility to Counter Poisoning in CTI LLM Pipelines
summary
The gist
Retrieval-Augmented Generation (RAG) systems in Cyber Threat Intelligence (CTI) are susceptible to poisoning attacks because sophisticated threat actors can mimic legitimate formats and terminology…
In short
RAGRank defends against poisoning attacks in Cyber Threat Intelligence (CTI) LLM pipelines by scoring source credibility using PageRank principles. It assigns authority scores to documents based on citation patterns and claim logic, then uses a two-pass retrieval strategy to prioritize trustworthy, high-authority context over potentially malicious or irrelevant information.
Key concepts
- RAGRank
- A novel defense mechanism that calculates a source credibility score for documents. It uses PageRank concepts adapted for CTI data to assign an authority value between 0 and 1 to each document based on how frequently it is cited, the commonality of its topics with other documents, and the logical consistency of its claims.
- Graph Construction Techniques
- Methods used to structure the document set into a network for scoring. These include using explicit citations (direct links), inferred citations (using an LLM to find topic overlap), and claim extraction (using an NLI model to check if one claim logically supports another). These structures define how authority flows between documents.
- Authority Scoring Mechanisms
- The process of calculating the final RAGRank score. This score combines a time-decayed PageRank rank, which penalizes older documents, with an author credibility factor derived from the average authority of their past work. This results in a comprehensive measure of a document's trustworthiness.
- Two-pass Retrieval Strategy
- A two-stage retrieval pipeline. First, it finds initial candidates using standard semantic similarity (cosine similarity). Second, it re-ranks these candidates using the calculated RAGRank score. This ensures that the final context fed to the LLM is not just semantically relevant but also comes from highly authoritative sources.
Terminology used across episodes
This episode discusses
- AuthRank: Scoring Source Credibility to Counter Poisoning in CTI LLM Pipelines · Paper Radio
- MS MARCO: A Human Generated MAchine Reading COmprehension Dataset
- From Local to Global: A Graph RAG Approach to Query-Focused Summarization
- ClaimTrust: Propagation Trust Scoring for RAG Systems
- ConfusedPilot: Confused Deputy Risks in RAG-based LLMs
- False Alarms, Real Damage: Adversarial Attacks Using LLM-based Models on Text-based Cyber Threat Intelligence Systems
- Towards More Robust Retrieval-Augmented Generation: Evaluating RAG Under Adversarial Poisoning Attacks
- Advancing Autonomous Incident Response: Leveraging LLMs and Cyber Threat Intelligence
- Certifiably Robust RAG against Retrieval Corruption
- TrustRAG: Enhancing Robustness and Trustworthiness in Retrieval-Augmented Generation
The paper
AuthRank: Scoring Source Credibility to Counter Poisoning in CTI LLM Pipelines · Read on arXiv
Austin Jia, Avaneesh Ramesh, Zain Shamsi, Daniel Zhang, Alex Liu
Applied Research Laboratories, The University of Texas at Austin, Texas, USA
Large Language Model (LLM) pipelines have become essential for processing Cyber Threat Intelligence (CTI) to quickly extract updated threat information from open source feeds and reports for real time processing. However, these pipelines are susceptible to poisoning attacks, and previously proposed defenses can fail for CTI contexts as cyber threat information is often completely novel, and attackers can easily mimic the stylistic conventions. We propose that the robustness of LLM pipelines can be improved by applying source credibility algorithms on document corpora, using PageRank as an example. We introduce AuthRank, a credibility score computed from citation networks built using explicit citations, LLM-inferred citations, and claim-level entailment, enhanced with time decay and author credibility. Our experiments on the standardized MS MARCO dataset demonstrate improved accuracy under poisoning, and we show proof-of-concept performance on CTI documents and feeds.
Transcript
Introduction to the show: ident: Security Radio. Generated commentary on the latest security and cryptography papers.
Nadia: We're looking at the paper: "AuthRank: Scoring Source Credibility to Counter Poisoning in CTI LLM Pipelines".
Elias: Here's the gist of the paper. Retrieval-Augmented Generation (RAG) systems in Cyber Threat Intelligence (CTI) are susceptible to poisoning attacks because sophisticated threat actors can mimic legitimate formats and terminology This paper proposes RAGRank, a novel defense that accelerates the robustness of modern RAG defenses by applying source credibility algorithms on corpora using PageRank as an example The algorithm applies a lower authority score to malicious documents while promoting trusted content when tested on the standardized MS MARCO dataset Problem Definition and Attack Model The adversary attempts to corrupt the document corpus D such that irrelevant and malicious information may be retrieved during the retrieval stage The attack succeeds when the retrieval set Rk for a query q contains a threshold number M of malicious documents dm, such that the LLM is forced to base its output y with significant consideration of the poisoned context In CTI contexts, there might be situations where Rk ≪ k for new information on emerging threats, providing an attacker a favorable opportunity to inject malicious information into threat feeds Graph Construction Techniques The methodology explores three ways to build a graph structure representing the data set: explicit citations, inferred citations, and claims extracted from the documents 1.
Nadia: Explicit Citations involve creating a directed citation graph based on each document’s explicit references 2.
Elias: Inferred Citations use an LLM to determine the commonality of topics between two documents as a weight for the connection in the citation network 3.
Nadia: Claim Extraction employs a pretrained NLI model to determine logical entailment between pairs of claims, treating each claim as a vertex in the graph Authority Scoring Mechanisms The final score RAGRank is calculated as the min-max normalized sum of time-decayed rank and author credibility The time-decayed authority score α'(di) is computed by scaling the initial PageRank score α(di) using a linear time-decay factor τ(di) based on the document's age Author credibility C(x) for an author x is defined as the average authority value of their prior documents Final Retrieval Strategy and Results The retrieval pipeline employs a two-pass ranking strategy where the top-k chunks are first retrieved based on cosine similarity and then re-ranked using their RAGRank score This two-stage approach ensures that only semantically related documents are considered for inclusion while promoting trustworthy and influential sources Experiments on the MS MARCO dataset show that RAGRank is able to improve accuracy of the LLM by about 10-15% in most cases In CTI proof-of-concept examples, RAGRank successfully ensures the LLM bases its answer on high authority documents even when poisoned chunks are highly similar to the query Future Work Directions Several avenues for extending this work have been identified, including performing more rigorous experimentation on larger CTI RAG databases with validation by domain experts Combining the three different graph building approaches into one technique could provide more robustness to the authority scoring Improving claim extraction by investigating the use of hierarchical summaries where many agreeing claims are grouped together before retrieval is suggested Exploring additional metadata options, such as follower/repost counts or domain type, could further build accurate source credibility Finally, studying possible attacks against RAGRank is necessary to counter long-term strategies where malicious actors build authority over time Table 1 TABLE 1.
Title and authors: Nadia: So we're looking at this paper called "AuthRank: Scoring Source Credibility to Counter Poisoning in CTI LLM Pipelines". It addresses a problem where Retrieval-Augmented Generation systems used in Cyber Threat Intelligence can get poisoned because bad actors can mimic legitimate threat formats and language.
Elias: Yeah, the authors are trying to make those RAG defenses more robust by using source credibility algorithms on the data itself, taking PageRank as their main example.
Nadia: They test this on the standardized MS MARCO dataset and they show that their algorithm gives a lower authority score to malicious documents while boosting trusted content.
Priya: I'm wondering what kind of malicious stuff they're talking about here, because in CTI, new threat info is super important, and if you down-rank something because it’s old or less cited, you might miss an emerging attack.
Nadia: That’s a fair question. The paper sets up the problem by saying that sometimes for new threats, the set of relevant documents might be very small compared to the total number of documents in the feed, giving attackers a chance to inject bad info where it matters most.
Elias: They build a citation network using three ways: explicit citations from the text, LLM-inferred citations based on topic commonality between two docs, and claim level entailment from an NLI model to see how claims logically connect.
Priya: So they're not just looking at where documents link to each other, but also if the actual claims within those documents actually support each other logically, which is a deeper look at the data quality.
Nadia: Exactly. And then they enhance that graph construction with refinements like time decay to fight against bias toward very recent stuff and author credibility based on an author's past work.
Elias: The final score they call RAGRank is a min-max normalized sum of a time decayed rank and the author credibility for each document, calculated using the initial PageRank score scaled by a linear time decay factor based on the document's age, plus the author credibility defined as the average authority value of that author's prior documents.
Nadia: And then they use this RAGRank score in a two-pass retrieval strategy: first you grab top k chunks using cosine similarity and then you re-rank those candidates using their RAGRank score to pick the final context for the LLM.
Priya: That two-stage process sounds like a good way to balance finding something semantically related with making sure it’s coming from a reliable source, but what happens if the initial cosine similarity pulls in a really bad document before the RAGRank re-ranking kicks in?
Elias: That’s where they show their results on the MS MARCO dataset. They found that this two-stage approach improves accuracy of answers by about ten to fifteen percent in most cases, and they also showed proof of concept performance on some CTI documents and feeds.
Title and authors: Nadia: It really shows how this mechanism can ensure the LLM bases its answer on high authority documents even when the poisoned chunks are very similar to what the user is asking for.
Priya: That's interesting because it suggests that sometimes prioritizing semantic relevance alone isn't enough, and you need that extra layer of source vetting to keep things accurate in a CTI setting.
Elias: The implication here is that we can actively defend the retrieval stage against poisoning by introducing this scoring mechanism.
Nadia: Looking ahead, the authors are thinking about a few things for future work. They want to do more rigorous experiments on bigger CTI RAG databases and have domain experts validate their findings.
Priya: And they’re also looking at combining the three different graph building approaches—explicit, inferred, and claim extraction—into one unified technique to see if that makes it even stronger against attacks.
Elias: They also want to investigate making claim extraction better by using hierarchical summaries where many agreeing claims are grouped together before the retrieval process starts.
Nadia: That makes sense for efficiency, but I think another area they point to is exploring more metadata like follower counts or domain types to build even more accurate source credibility scores.
Priya: And I have to ask, what about the attacks against AuthRank itself? They don't seem to cover long-term strategies where malicious actors build up authority over time, which could bypass these initial scoring methods.
Elias: That’s a valid concern. The authors themselves flagged that studying possible attacks against their RAGRank system is necessary to counter those kinds of long-term poisoning tactics.
Nadia: So, we have this defense that uses graph theory and PageRank to score sources, a two-pass retrieval strategy, and it shows improvements on standardized data sets like MS MARCO.
Priya: It’s a solid step forward in making CTI systems less vulnerable to corpus poisoning by adding this layer of structural credibility assessment.
Elias: Ultimately, the paper "AuthRank: Scoring Source Credibility to Counter Poisoning in CTI LLM Pipelines" shows how applying source credibility algorithms via PageRank can accelerate the robustness of RAG defenses against poisoning attacks.
Nadia: It gives us a concrete way to filter context before it even gets to the LLM, which is crucial for high-stakes fields like threat intelligence.
Priya: For anyone using these systems, it means you can’t just trust that the most relevant document found by a search engine is actually trustworthy; you need this kind of scoring.
Elias: The authors show that this method works well, improving accuracy by ten to fifteen percent on the MS MARCO dataset and providing a proof of concept in CTI documents.
Nadia: It’s a practical defense mechanism that focuses on establishing authority based on how documents reference each other and how those claims align logically.
Priya: The future work points toward making this even more comprehensive by integrating different data construction methods and looking at temporal dynamics to keep up with evolving threats.
The paper's summary: Nadia: So, we're looking at "AuthRank: Scoring Source Credibility to Counter Poisoning in CTI LLM Pipelines." Basically, they’ve built a system to score the trustworthiness of documents feeding into those RAG systems used for Cyber Threat Intelligence.
Elias: Right. The core idea is applying source credibility algorithms, specifically PageRank, to the document corpus itself. They assign a score between zero and one to every single document based on how well it cites others and how reputable the authors are over time.
Priya: So instead of just taking the top twenty documents that look semantically similar, they’re filtering those twenty using this authority score. That means a relevant but low-quality source might get filtered out in favor of a slightly less relevant but much more authoritative one.
Nadia: Exactly. They use three ways to build this graph structure: explicit citations from the text, inferred connections where an AI figures out which documents talk about the same topics, and claim extraction where they check if the actual claims within documents logically support each other.
Elias: And they combine those with time decay—so newer information gets a boost initially—and author credibility, which is just an average of all the authority scores of that author's previous work.
Priya: What’s interesting is how they handle the retrieval part too. It’s a two-pass system: first you find chunks using simple cosine similarity, and then you re-rank those candidates using this calculated RAGRank score before feeding them to the LLM.
Nadia: And they showed that when tested on the MS MARCO dataset, this method improves the final answer accuracy by about ten to fifteen percent in most scenarios. That’s a solid bump in performance for a system dealing with noisy intelligence feeds.
Elias: It’s designed to specifically counter poisoning attacks where bad actors try to inject malicious info into threat feeds. If they can't get that document high enough on the authority score, it gets demoted, even if it looks relevant to the query.
Priya: But I gotta ask, how does this handle those tricky situations in CTI where new threats emerge? Like if there isn't much existing literature yet, a low authority score might unfairly suppress the very new information we need to see.
Nadia: That’s a real concern for them. They actually included a proof of concept showing that when an attacker tries to front-run the system with fake info about a brand new domain, AuthRank correctly flags that there isn't enough authoritative information about it yet.
Elias: So they're not just looking at existing reputation; they’re modeling the structure of knowledge itself through citations and logical claims. It changes how we think about what constitutes reliable data in these massive, messy threat databases.
Priya: It suggests that for CTI, relevance isn't enough; you need structural trust built into the retrieval pipeline itself to handle both existing knowledge and emerging threats fairly.
The paper's improvements: Tom: We’re looking at how they suggest improving their original AuthRank system to make it even tougher against these poisoning attempts.
Nadia: They’ve got several future work directions that focus on making the scoring more comprehensive and robust across different data sources.
Elias: Yeah, one big idea is combining those three graph building approaches—explicit citations, inferred citations, and claim extraction—into one unified technique.
Priya: That makes sense for depth; if you use just one way to score authority, you’re missing the full picture of how information actually connects.
Nadia: Exactly. They also look at making the claim extraction better by using hierarchical summaries where they group many agreeing claims together before suggesting retrieval.
Elias: That sounds like a smart way to reduce noise; if many different documents all agree on one piece of information, that single claim gets a stronger signal for scoring.
Priya: And they’re also thinking about adding more metadata options, like follower or repost counts or domain type, to build even better source credibility scores. That moves beyond just internal citations.
Nadia: It’s about giving the system more external context on how influential a source is in the broader threat landscape.
Elias: But I gotta point out that they also flag studying attacks against AuthRank itself as necessary for long-term strategies where attackers build up authority over time, which could bypass these initial scoring methods.
Priya: That’s a good caution; they aren't claiming this is a permanent fix against every kind of poisoning strategy yet. They admit the system needs to be tested against those slower, persistent attacks.
Nadia: So, the implication for us in practice is that we need to keep watching how these authority scores evolve over time as threat actors adapt their poisoning methods.
Elias: It means this isn't a set-it-and-forget-it solution; it’s an ongoing research area for defense. We still have to figure out how to model those adversarial tactics that build up influence slowly.
Conclusion: Nadia: So we’re wrapping up our look at "AuthRank: Scoring Source Credibility to Counter Poisoning in CTI LLM Pipelines." Essentially, this work shows how you can build a structural defense into Retrieval-Augmented Generation systems by scoring document authority using graph algorithms like PageRank.
Elias: Right. It moves the defense from just looking at the documents themselves to analyzing the relationships between them—citations and claims—to assign a credibility score that fights poisoning attacks.
Priya: What this really means for us is that when we use these systems in CTI, we can trust a retrieval set more because it’s filtered not just by what sounds relevant, but by how much structural support the document has from other reliable sources.
Nadia: It gives us a tangible way to see if a document is merely semantically similar to our query or if it’s actually built on solid, traceable knowledge.
Elias: And the results on MS MARCO show that this two-pass ranking strategy, where you re-rank based on that RAGRank score after the initial similarity search, actually boosts accuracy by ten to fifteen percent in most cases.
Priya: But I still have to ask about those limitations mentioned by the authors. They flag that this method doesn't account for long-term poisoning strategies where attackers build up authority over months or years.
Nadia: Exactly. It’s a defense against current, immediate attacks and front-running, but it doesn't cover the slow erosion of trust over time.
Elias: So the implication is that we need to keep studying how these authority scores change dynamically; it’s not a static solution for everything.
Priya: It also highlights that for systems handling sensitive data like threat intelligence, relevance alone isn't enough; you need this kind of deep structural vetting baked into the pipeline.
Nadia: For anyone listening, keep an eye on how these authority scores behave as you use these RAG systems in your own work, because understanding those numbers is key to spotting manipulation.
Elias: This paper on AuthRank gives us a solid starting point for defense, but we’ll need more research into those long-term adversarial models to make this truly bulletproof.
More episodes
- 2608.27496-ROPE: Routed Origin Policy Enforcement against Indirect Prompt Injection
- Daily Summary for 2026-11-19
- 2607.07085-Is Randomness Necessary for Adaptive Data Analysis?
- The day in conversation for 2026-11-19
- Daily Summary for 2026-11-19
- 2609.10601-Threshold Choice, Not Sample Size, Bounds Trustless Verification of Nondeterministic Compound AI Workflows
- 2610.38786-Adversarial Process Disturbances at a Deployable Alarm Budget: Detector Evaluation on a Simulator and a Real Plant
- 2511.08660-Binary and Multiclass Cyberattack Classification on GeNIS Dataset
- 2604.27302-Static Attribution of Android Residential Proxy Malware Using Graph Kernels
- Daily Summary for 2026-11-18