A systematic analysis of indirect prompt injection through tool call responses in LangChain, LlamaIndex and AutoGen-style agents — how malicious content in external data sources can hijack agent behaviour and the controls that mitigate it.
Recent research demonstrates that vision-language models including GPT-4V, Gemini Pro Vision, and open-source alternatives are highly susceptible to adversarial image perturbations, with attacks transferring across models at rates significantly higher than classical vision model attacks.
Deepfake video and audio fraud against financial institutions reached record levels in Q1 2026, driven by the commoditisation of real-time face-swap and voice cloning tools now available for under $50/month on criminal markets.
A survey of query-efficient model extraction attacks against commercial LLM APIs — how adversaries can reconstruct a functional shadow model using only input-output pairs, the commercial and security risks this creates, and the defences providers are deploying.