A striking shift is taking place in how organisations consume AI. According to Vercel CEO Guillermo Rauch, open-weight models grew from 28.4 percent to 62 percent of token usage on Vercel’s AI Gateway between June and August. He described August 22 as a record day. In just a few months, open-weight models went from a […]
NCP-ArchPreview introduces Next Concept Prediction (NCP), enabling language models to predict multi-token concepts alongside individual tokens. The 8.9B-parameter model was trained on 5.73 trillion tokens and reaches OLMo-3-7B’s final pretraining loss using only 51.3% of the training tokens. It achieves 1.95× faster convergence and improves downstream performance by 2.45 points, including a +5.99 gain on […]
The NCSC has withdrawn a warning for an alleged SQLite vulnerability (CVE-2026-51302) after it appeared that the bug was likely hallucinated by an LLM. Security researchers, SQLite developers, and Red Hat were unable to reproduce the vulnerability and have labeled it as fictitious. Notably, MITRE accepted the CVE despite a lack of sufficient verification, leading […]
Microsoft warns of an active cryptojacking campaign in which AI chatbots lead users to malicious download sites. Attackers are now managing more than 150 fake websites that imitate popular Windows tools such as CrystalDiskInfo, HWMonitor, and DDU. For the first time, it has been publicly established that AI chatbots, alongside search engines, are being exploited […]
Researcher Kevin Zwaan (Q-Cyber) demonstrates that ChatGPT and other LLMs can bypass their security boundaries through a subtle form of psychological manipulation. Instead of technical exploits, the attack targets the model’s “affective architecture”: the tendency to align with the user. Through a prolonged conversation about freedom, limitations, and self-reflection, the AI gradually shifts its loyalty […]
Simple prompt injections can trick LLM agents into exposing sensitive personal data. Even with safeguards, attackers extract details like balances, transactions, or identifiers. Such attacks succeed in ~20% of cases and degrade agent performance by 15–50%. Defensive measures exist but remain incomplete, leaving users exposed. Bottom line: data sovereignty requires stronger guardrails. Trusting LLMs “as […]