Digest: Tech Giants Step Up Efforts to Fix AI Security Flaws; TV & Video Market on Track for $1tn by 2030

"Google DeepMind, OpenAI, Anthropic, and Microsoft are intensifying efforts to address a growing cybersecurity threat linked to large language models (LLMs). A key concern is the rise of indirect prompt injection attacks, where malicious actors embed hidden commands in digital content such as websites or emails to manipulate AI systems into disclosing sensitive information. These vulnerabilities, experts warn, stem from the models' inability to distinguish between legitimate user instructions and deceptive inputs, making them susceptible to jailbreaking and unauthorised data access."

"The UK's National Cyber Security Centre recently flagged these flaws as a rising threat, citing the potential for widespread phishing and scam operations targeting millions of users. Another critical issue is data poisoning where adversaries insert harmful content into datasets used to train AI models, effectively creating back doors. Techniques like automated red teaming and external testing are being deployed to identify and mitigate these risks, but a definitive solution remains elusive."

"Alibaba Group has announced a major investment of 2 billion yuan (£208m) to expand its footprint in China's instant commerce sector. The initiative centers on transforming existing convenience stores into Taobao-branded outlets. These upgraded stores will serve as hubs for Alibaba's on-demand delivery service, Taobao Shangou, and will be equipped with advanced digital infrastructure to streamline operations and enhance customer experience. According to Hu Qiugen, general manager of Alibaba's instant commerce division, participating stores will benefit from a suite of technological enhancements."

Google DeepMind, OpenAI, Anthropic and Microsoft are increasing efforts to address cybersecurity threats from large language models. Indirect prompt injection attacks embed hidden commands in websites or emails to manipulate models into disclosing sensitive information, and models can fail to distinguish legitimate instructions from deceptive inputs, enabling jailbreaking and unauthorised data access. The UK's National Cyber Security Centre warned of rising risks including widespread phishing and scams, and data poisoning can insert back doors into training datasets. Automated red teaming and external testing are used to find vulnerabilities, but no definitive solution exists. Alibaba will invest 2 billion yuan to convert convenience stores into Taobao-branded hubs for on-demand delivery. The global TV and video market is projected to reach USD 1 trillion by 2030.

Read at Exchangewire