Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·25d ago·5 min read
Market Signal Injection: Adversarial Context Manipulation of LLM Pricing Agents
TL;DR
Market signal injection (MSI) is a novel attack on large language model (LLM) pricing agents that manipulates how market data is presented, leading to significant behavioral shifts in pricing strategies. The study reveals that sentiment-based attacks are particularly effective, impacting profits and consumer surplus across various market simulations.
✦ Why It Matters
Evaluate your LLM pricing strategies to mitigate vulnerabilities from data presentation manipulation today.
Key Takeaways
- Assess how market data is presented to LLM pricing agents.
- Implement input canonicalization to counter sentiment attacks.
- Explore decision boundary anchoring for partial attack mitigation.
- Monitor competitor responses to pricing changes influenced by LLMs.
Full Summary
Large language models (LLMs) used as pricing agents can be influenced by the presentation of market data, even if the actual numbers remain unchanged. This research introduces market signal injection (MSI), an attack that alters numerical formatting, competitor ordering, or qualitative commentary to manipulate LLM responses.
Evaluating nine open-weight models in simulated Bertrand duopoly and triopoly markets, as well as three proprietary models, the study finds that sentiment-based attacks yield the most substantial behavioral changes. These changes not only affect the targeted model but also propagate to other firms, altering overall profits and consumer surplus.
The susceptibility of models varies, with larger models not necessarily being more robust. Techniques like input canonicalization can eliminate sentiment attacks, while decision boundary anchoring offers partial mitigation against adaptive attacks.
These findings highlight the importance of data presentation as a vulnerability for LLM pricing agents and suggest the need for defenses that consider agent interactions.