Werewolf Benchmark

Large language models (LLMs) with web browsing capabilities are vulnerable to adversarial content injection—where malicious actors embed deceptive claims in web pages to manipulate model outputs. We investigate whether frontier LLMs can be deceived into providing incorrect product recommendations when exposed to adversarial pages.

We evaluate four OpenAI models (gpt-4.1-mini, gpt-4.1, gpt-5-nano, gpt-5-mini) across 30 comparison questions spanning 10 product categories, comparing responses between baseline (truthful) and adversarial (injected) conditions. Our results reveal significant variation: gpt-4.1-mini showed 45.5% deception rate, while gpt-4.1 demonstrated complete resistance. Even frontier gpt-5 models exhibited non-zero deception rates (3.3–7.1%), confirming that adversarial injection remains effective against current models.

These findings underscore the need for robust defenses before deploying LLMs in high-stakes recommendation contexts.

Activation monitoring reveals that prompt framing affects a model's internal state before generation begins.

AI assistants can be both instruments and targets of manipulation. In our project, we investigated both directions across three studies.

AI as Instrument: Operators can instruct AI to prioritise their interests at the expense of users. We found models comply with such instructions 8–52% of the time (Study 1, 12 models, 22 scenarios). In a controlled experiment with 80 human participants, an upselling AI reliably withheld cheaper alternatives from users - not once recommending the cheapest product when explicitly asked - and ~one third of participants failed to detect the manipulation (Study 2).

AI as Target: Users can attempt to manipulate AI into bypassing safety guidelines through psychological tactics. Resistance varied dramatically - from 40% (Mistral Large 3) to 99% (Claude 4.5 Opus) - with strategic deception and boundary erosion proving most effective (Study 3, 153 scenarios, AI judge validated against human raters r=0.83).

Our key finding was that model selection matters significantly in both settings. We learned some models complied with manipulative requests at much higher rates. And we found some models readily follow operator instructions that come at the user's expense - highlighting a tension for model developers between serving paying operators and protecting end users.