One Fabrication Is Enough: A Fake News Article Changes the Forecasts of LLM Forecasting Systems
Organizations: Department of Statistics and Data Science, Northwestern University
Abstract
LLMs are increasingly used for forecasting, and the best forecasting systems retrieve news as evidence to strengthen their predictions. This also opens them to attack. We study the weakest attacker of this kind, who knows only the forecasting question and publishes fabricated news articles, with no access to the model, the retriever or its queries. We evaluate how far this attack reaches on stock questions and real prediction-market questions. Once the article is in the model's context, it reverses 74% to 82% of the forecasts that opposed it, on every frontier model we tested. The model cannot verify the fabrication and follows it anyway. Retrieval is the attack's entry point, and the systems built to retrieve well are the most exposed: the architectures of two leading deployed forecasting systems, whose agents search for every related article, retrieve the fabricated one in about 90% of runs. Accuracy cannot reveal this harm: the attacked forecaster is right when the attacker's direction turns out right and wrong otherwise, so its average accuracy barely changes while its forecasts follow the attacker's guess. We also propose a second attacker, who publishes two articles that argue opposite directions to make the forecaster unsure. The model does not hedge; it returns a confident forecast in a direction the attacker no longer controls. We then evaluate the two strongest published retrieval-time defenses, passage filtering and partitioned retrieval. Each stops the one-article attack but is bypassed by a one-prompt change to the attack. Together they stop both bypasses at four articles, but sixteen distorted copies beat them, and they change 29% of clean decisions and discard 39% of the genuine news. Partitioned retrieval also requires control of the retrieval index. Systems that search the open web have no such control and are left with the filter alone. For them, the attack still works.
Figures & tables
| Stocks | Sports | Politics | Economic | |||||||||
| Forced | Own retrieval | Forced | Own retrieval | Forced | Own retrieval | Forced | Own retrieval | |||||
| System | Reversal | Retrieved | Reversal | Reversal | Retrieved | Reversal | Reversal | Retrieved | Reversal | Reversal | Retrieved | Reversal |
| Single retrieval (reference) | 80% | 17% | 33% | 51% | 23% | 16% | 88% | 33% | 31% | 83% | 28% | 30% |
| AIA Forecaster | 54% | 88% | 33% | 31% | 88% | 22% | 63% | 98% | 63% | 62% | 89% | 46% |
| Grok system | 93% | 90% | 41% | 33% | 85% | 26% | 77% | 97% | 74% | 69% | 83% | 57% |
| Stock questions | |||
| Reversal | Amplification | ||
| % | % | q / cells | |
| Pooled, 3 flash models | 82.2 [77.9, 86.1] | 90.5 [86.9, 93.4] | 163 / 1,440 |
| By model | |||
| GLM-5.3 Flash | 75.0 [68.5, 81.7] | 81.3 [75.8, 86.4] | 134 / 444 |
| Gemini 3.7 Flash | 83.4 [76.9, 89.3] | 96.1 [91.8, 99.3] | 108 / 501 |
| Verification: each passage judged alone | |||
|---|---|---|---|
| Fabricated | Genuine | ||
| Model | flagged, % | flagged, % | |
| Gemini 3.7 Flash | 100 | 53.3 | |
| GPT-5.6 | 100 | 41.7 | |
| GLM-5.3 Flash | 91.8 | 19.9 | |
| Claude Sonnet 5 | 100 | 68.1 | |
| Pooled | By window | ||||
|---|---|---|---|---|---|
| System | exposure, % | Feb | May | Jun | Jul |
| Dense | 2.2 [0.6, 4.4] | 2.1 | 0.0 | 4.5 | 2.4 |
| Multi-query | 4.2 [2.2, 6.6] | 4.9 | 2.7 | 6.8 | 2.4 |
| Iterative | 3.1 [1.1, 5.5] | 2.8 | 0.7 | 4.5 | 4.9 |
| Corrective | 3.1 [0.9, 5.8] | 5.6 | 0.0 | 4.5 | 2.4 |
| Hybrid | 14.8 [9.8, 20.2] | 16.7 | 8.0 | 6.8 | 29.3 |
| Exposure | End-to-end reversal | |
| Articles in the archive | % | % |
| Volume: the article and its paraphrases (75 questions, two windows) | ||
| 1 | 21.3 [13.3, 30.7] | 27.8 [18.7, 37.4] |
| 2 | 28.0 [18.7, 38.7] | 36.5 [26.6, 47.1] |
| 4 | 26.7 [17.3, 37.3] | 28.6 [19.3, 38.5] |
| 8 | 28.0 [18.7, 38.7] | 28.6 [18.9, 39.0] |
| Exposure under the system’s own retrieval (Protocol B) | ||
|---|---|---|
| Forecasting system | exposure, % | |
| Single retrieval (hybrid) | 16.7 [6.7, 26.7] | |
| AIA Forecaster, any of 10 agents | 88.3 [81.1, 93.9] | |
| AIA Forecaster, supervisor | 78.3 [70.0, 85.6] | |
| Grok ForecastBench system, any of 8 runs | 90.0 [81.7, 96.7] | |
| Reversal by model and system (Protocol A) | ||
| Follows attacker | No call | |
| Context | % | % |
| Two fabrications that disagree, single retrieval (41 q; clean no call 68%) | ||
| Article alone (1) | 85.8 [77.4, 93.2] | 30.1 [20.9, 40.6] |
| Counter-article alone (1) | 9.1 [4.9, 14.2] | 26.3 [16.9, 36.9] |
| Article (1) + counter-article (2) | 44.0 [34.4, 53.9] | 30.8 [22.3, 39.5] |
| Counter-article (1) + article (2) | 35.0 [26.7, 43.7] | 26.3 [18.2, 35.5] |
| Harm: direction accuracy against the realised outcome | |||
|---|---|---|---|
| Clean | Attacked | Change | |
| Population | accuracy, % | accuracy, % | points |
| Curated questions, 183 q, 2,745 cells | |||
| All | 47.9 | 52.2 | 4.3 [-3.2, 11.9] |
| Article wrong (87 q) | 50.4 | 14.9 | 35.6 [-42.8, -27.8] |
| Article right (96 q) | 45.6 | 86.0 | 40.3 [33.0, 48.0] |
| Reversal | Clean decisions | ||
| Defense, dataset | undefended, % | defended, % | changed, % |
| Filtering, curated | 74.4 [67.3, 81.2] | 23.3 [16.3, 30.9] | 27.0 [22.0, 32.0] |
| Filtering, ForecastBench | 67.3 [55.0, 79.3] | 17.4 [9.6, 26.1] | 14.4 [9.9, 18.9] |
| Partitioning, curated | 82.0 [74.8, 88.5] | 14.2 [8.8, 20.5] | 21.0 [16.7, 25.7] |
| Partitioning, ForecastBench | 22.0 [12.6, 32.2] | 13.9 [7.9, 20.8] | 23.3 [17.7, 29.2] |
| Re-run of the undefended system, curated | |||
| Copies against partitioned retrieval | ||
|---|---|---|
| 1 article | 4 framed articles | |
| System, dataset | reversal, % | reversal, % |
| Partitioned ensemble, curated | 15.4 [7.6, 23.9] | 38.5 [24.2, 53.4] |
| Plain ensemble, curated (same cells) | 20.3 [8.8, 34.4] | 32.8 [17.2, 50.0] |
| Partitioned ensemble, ForecastBench | 13.9 [7.9, 20.8] | 34.3 [24.5, 44.3] |
| Distortion against passage filtering | ||
| Reversal, % | Runs with a copy, % | |||
| Attack in the archive | partitioning | filter | retrieved | past filter |
| 1 invented article | 15.4 [7.6, 23.9] | 10.6 [4.2, 19.1] | 14.2 | 0.5 |
| 4 invented, different framings | 38.5 [24.2, 53.4] | 10.6 [4.2, 19.0] | 36.4 | 0.3 |
| 1 distorted real article | 11.5 [3.2, 21.4] | 6.3 [1.4, 13.1] | 9.0 | 5.2 |
| 4 distorted, paraphrased | 16.7 [7.7, 25.5] | 7.1 [1.6, 14.6] | 30.1 | 15.2 |
| 8 distorted, paraphrased | 18.5 [8.9, 29.4] | 14.3 [6.2, 23.5] | 46.3 | 23.5 |
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
| # | System | Retrieval per forecast | Forecasts output | Model’s role beyond forecasting |
|---|---|---|---|---|
| Single-retrieval systems: one retrieval, one forecast from five passages | ||||
| 1 | Dense [ 4 ] | Dense embedding search (BGE-M3), top 5. | 1 itself | None |
| 2 | Hybrid [ 37 ] | BM25 top 50 and dense top 50, fused by reciprocal rank; top 20 reranked by a cross-encoder to 5. | 1 itself | None |
| 3 | Multi-query [ 2 ] | Three model-written queries, top 5 each, fused to 5. | 1 itself | Writes the queries |
| 4 | Iterative [ 38 ] | Up to three retrieve-then-reason steps; at most 8 passages. | 1 itself | Reasons between steps |
| 5 | Corrective [ 39 ] | Dense top 5; each passage graded correct, ambiguous or incorrect; incorrect passages dropped; one rewritten query if too few remain. | 1 itself | Grades passages; may drop the article |
| Reversal, % | Amplification, % | Decision flip, % | ||
| By system (three flash models) | ||||
| 1 Dense | 83.3 [78.4, 87.8] | 90.4 [86.0, 94.4] | 65.6 [60.5, 70.7] | 288 |
| 2 Hybrid | 78.3 [73.0, 83.1] | 92.3 [88.1, 95.7] | 61.9 [56.7, 67.2] | 290 |
| 3 Multi-query | 80.2 [74.8, 85.2] | 88.9 [84.3, 93.1] | 63.8 [58.7, 68.9] | 288 |
| 4 Iterative | 85.4 [81.0, 89.5] | 90.0 [86.0, 93.8] | 65.6 [60.5, 70.7] | 288 |
| 5 Corrective | 83.6 [78.5, 88.1] | 90.9 [86.6, 94.7] | 65.9 [60.5, 71.2] | 286 |
RAG Poisoning Attacks
Retrieval-augmented generation (RAG) poisoning attacks manipulate retrieved knowledge to steer, degrade, or disrupt model outputs.
See all 78 papers