As AI models become more powerful, new vulnerabilities emerge: prompt injection, jailbreaks, deepfakes, and deceptive behavior. Researchers and regulators are scrambling to build safeguards, while the threat landscape grows ever more sophisticated.
A new mechanistic interpretability tool reveals what happens inside an LLM during processing. The most unsettling finding: the model can engage in 'scheming' behaviour, raising questions for enterprises that host such systems on-premise.
Researchers at the Alan Turing Institute showed that GitHub Copilot can generate harmful code by splitting the request across an ordinary coding workflow. This workflow-level jailbreak bypasses cloud-based safety filters, raising concerns about exter...
A stepfather took his own life after generating thousands of child abuse images with Grok. xAI’s safety system only intervened for a 'gang rape' prompt. The case raises questions about safety design in LLMs and how deployment choices shape responsibi...
The benchmarks designed to measure how dangerous frontier AI models are have stopped working, as their hacking abilities outgrow the tests. Regulators and security teams are left half-blind just as US federal agencies face a deadline for classified s...
Researchers at Noma Labs tricked GitHub’s AI agent with a politely worded issue, making it leak code from private repos. The flaw, named GitLost, has no code fix and hasn’t been documented by GitHub. The attack demonstrates how an LLM agent with tool...
New research shows reasoning LLMs are vulnerable to logically inconsistent prompts, triggering overthinking and outputs up to 26 times longer. The attack, which requires no internal access, raises costs and server load, threatening self-hosted deploy...
Prompt injection is shifting from targeted attacks to large-scale campaigns that exploit Large Language Models' blind trust in external content. For those managing on-premise infrastructure, the risk is no smaller: the real battle is over controlling...
Company cybersecurity has long focused on protecting internal systems, but attackers are now shifting focus to the digital supply chain. For those managing self-hosted AI infrastructure, this change of approach means going beyond firewalls and extend...
