Giskard in the news

Giskard offers a platform with advanced red teaming agents to test and detect security vulnerabilities in AI agents used by major French companies, focusing on multi-turn adversarial attacks and data exfiltration risks.

In July 2025, Giskard's Phare LLM benchmark ranks Meta's Llama and Google's Gemini models as most reliable in resisting hallucinations and harmful behaviors, while some models like X's Grok 2 perform poorly.

Giskard's study finds that large language models generate biased content yet inconsistently recognize these biases, highlighting a gap between bias generation and detection due to fine-tuning practices.

Giskard's Phare benchmark reveals that major language models frequently hallucinate, influenced by query tone and brevity, with popular models showing significant error rates.

Giskard's Phare project benchmarks major LLMs on hallucination, bias, and harmful content risks, revealing that popular models like GPT-4o are not the most factual or robust.

Giskard's study reveals that AI chatbots often produce more factual errors when prompted to give short, concise answers, affecting even top models like GPT-4o.

Giskard's analysis shows that certain phrases and short response requests increase AI hallucinations.

Giskard's study shows that language models hallucinate more with concise or confident prompts, reducing factual accuracy and correction ability.

A study reveals that concise chatbot responses significantly increase hallucination rates.

Instructing AI models to provide short answers often leads to more confident but incorrect responses, reducing their factual reliability.

Giskard's study reveals that AI models produce more factual errors when forced to give brief answers, as brevity limits nuanced explanations and error correction.

Short prompts to AI chatbots increase hallucinations by limiting context, causing more false information in responses.

Telling AI chatbots to give short answers increases hallucinations by limiting their ability to correct false information.

Giskard found that AI chatbots hallucinate more when prompted to give short answers, as brevity can compromise accuracy.

Giskard's new benchmark reveals which AI models are most prone to hallucinations.

Giskard's new benchmark Phare evaluates the risks of AI language models, highlighting the worst performers.

Giskard's president talks about their AI antivirus solution on a business podcast.

Giskard's president introduced an antivirus for AI agents on a business show.

Giskard specializes in testing AI systems to ensure quality and security, particularly for generative AI applications.