Fabricated news can steer LLM trading agents into the same trades
Original title: Fake News, Real Crashes: Information Injection Attacks on LLM Financial Trading Agents
On about 10,000 decisions across 15 U.S. stocks, two 30-word fake-news templates hit 100% attack success.
A benchmark study finds LLM trading agents that act on news can be pushed into synchronized trades by fabricated evidence rather than altered instructions. Across five injection vectors and seven event categories, fabricated news reached 62% attack success versus 14% for prompt injection. Surface cues such as a named source or a "BREAKING" tag drove fact-poisoning success, and most attacks still worked against standard defenses.
Unconfirmed: Success rates are the authors' benchmark results, not confirmed losses in live markets.
- Two 30-word templates reached 100% success on all 15 stocks.
- Dropping a named source or "BREAKING" tag cut that rate from 100% to 0%.
- A fake indicator panel scored 91% against an agent about to hold and 12% against one about to buy.
- Attacks generalized across model vendors and worked when blended unlabeled into real feeds.
- Scaling and peer debate failed; only hierarchical aggregation reduced success.
- Countries
- United States
From the original article
1 min read at arxiv.orgLarge language model (LLM) agents now trade real capital using news as evidence, yet they cannot reliably tell fabricated headlines from real news. Because they read the same feeds under near-identical policies, a single fabricated post can push them into the same trade at the same moment. We present a systematic attack benchmark for information injection against LLM trading agents.
By Mengxiao Wang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.