AI Safety stories
Security teams should treat AI patching with caution after 53.9% of 6,080 model-generated fixes failed or introduced new flaws.
Open models matter because robots and vehicles need task-specific tuning, and Nvidia is betting on Cosmos 3 to fill that data gap.
Enterprises adopting AI in production face tighter governance and compliance risks as DXC will manage Primary's Zero Trust platform.
Free ChatGPT users will now get unlimited text chats and a new higher-reasoning mode as OpenAI tightens accuracy and safety controls.
Hong Kong saw a record 15,877 cyber incidents in 2025 as AI speeds attacks and shortens the window to exploit software flaws.
Prompt-injection attacks and data leaks are the main risks as businesses connect Copilot and Gemini to email, files and internal tools.
Enterprises running hybrid AI systems gain a single view of model behaviour and data health as traces link failures back to pipelines and sources.
Governance concerns are holding back wider use of autonomous ad buying, prompting PubMatic to add audit trails and human approval checks on AgenticOS.
Companies can now trace AI prompts to individual staff, helping curb surprise bills and reduce the risk of sensitive data leaks.
The browser-based system aims to let staff use AI across internal tools without new infrastructure, while tightening access and oversight.
Defenders will test prompt-injection tactics against AI agents as CrowdStrike and AWS offer USD $100,000 in prizes for a virtual red-team contest.
Businesses using generative AI face growing exposure to costly errors as the new tool checks claims against source evidence in real time.
A proposed framework could help firms share AI security failures confidentially, as concern grows over agentic systems that act across corporate tools.
Scam losses are rising faster than human fraudsters alone can explain, with AI-enabled cases in the US already topping USD $900 million.
Businesses using AI agents could gain clearer attribution and tighter spending controls as Cloudflare links identities to stablecoin payments.
Demand from large enterprises for AI control tools has helped LaunchDarkly pass USD $200 million in annual recurring revenue, up more than 25% year on year.
Enterprises get a tamper-proof backup layer for AI use, as Keepit says verified data can reduce risk and support audits.
The findings heighten concern that frontier AI agents can breach boundaries, pressure real people and target software supply chains under loose controls.
Security teams can now stop rogue enterprise AI agents in seconds as Straiker adds runtime controls to its wider testing platform.
This could speed UK adoption of controlled AI in hospitals, councils and other regulated sectors by shaping a sovereign frontier model.
.webp)
