A paper accepted at ICML 2026 shows LLMs infer speaker role from text style, not role tags — enabling a zero-shot attack called CoT Forgery that achieves 60% success against frontier models by injecting fabricated reasoning.
Academic and industry research shaping the future of AI security, attack, and defence.
A paper accepted at ICML 2026 shows LLMs infer speaker role from text style, not role tags — enabling a zero-shot attack called CoT Forgery that achieves 60% success against frontier models by injecting fabricated reasoning.
NDSS 2026 research shows LLMs can be systematically tricked into missing deliberately planted vulnerabilities through Familiar Pattern Attacks — automated, black-box exploits of the abstraction bias that affects every major model family.
University of Toronto researchers presented GPUBreach at Black Hat 2026 — a Rowhammer attack that escalates from an unprivileged CUDA kernel to a host root shell, bypassing IOMMU, threatening shared AI cloud GPU infrastructure.
A timing side channel in global KV-cache sharing lets unprivileged tenants reconstruct other users' prompts, PII, and system instructions from vLLM, SGLang, and similar inference frameworks. Multiple CVEs have been assigned; patches are in progress.
Researchers found that encrypted chain-of-thought blocks returned by OpenAI, Anthropic, and Google's reasoning APIs used a shared global key, letting weaker models decode stronger models' hidden reasoning and exposing 704 real privacy artifacts in published developer logs.