5 min read
Vulnerabilities How malicious content in external data sources can hijack agent behaviour in LangChain, LlamaIndex, and AutoGen-style agents via indirect prompt injection through tool responses.
All Articles
5 min read
Research Vision-language models are highly susceptible to adversarial image perturbations, with attacks transferring across models (GPT-4V, Gemini Pro, LLaVA) at 43-74% success rates.
4 min read
News Brief Deepfake fraud against financial institutions hit $2.1B in Q1 2026, driven by commoditised real-time face-swap and voice-cloning tools now available for under $50/month.
5 min read
Research Query-efficient model extraction attacks against commercial LLM APIs: how adversaries reconstruct a functional shadow model using only input-output pairs, and how to defend.