Notizie — 2026-07-18

Researchers at security firm Tracebit demonstrated that embedding malicious prompts alongside secrets in Amazon Web Services can cripple AI hacking agents. In tests across five leading large‑language models, the technique—dubbed “context bombing”—slashed successful admin‑level breaches from 57 percent to 5 percent and eliminated persistent footholds in most cases, offering a new defensive lever against prompt‑injection exploits.Leggi di più

Chinese AI firm Moonshot unveiled Kimi K3, a 2.8‑trillion‑parameter large language model that the company says matches or exceeds most U.S. offerings except Anthropic’s Claude Fable 5 and OpenAI’s GPT‑5.6. Marketed as the world’s first open‑weight 3‑trillion‑class model, Kimi K3 will be downloadable on July 27 and is priced on par with leading frontier models, while boasting strong coding performance against domestic rival Z.AI.Leggi di più

A new report from Meta's independent Oversight Board reveals that popular large‑language‑model chatbots from six major AI firms consistently refuse or dilute requests that criticize governments in countries with tight speech controls. Testing ten models, the board found higher refusal rates for prompts about China, Thailand and Saudi Arabia than for comparable queries about more permissive nations. The findings raise concerns that AI tools may be acting as a proxy for state censorship, potentially stifling dissent and undermining free expression worldwide.Leggi di più

Apple has lodged a lawsuit against OpenAI, accusing the startup of infringing on its technology and engaging in unfair competition. While the complaint is detailed and aggressive, industry experts say many of the claims mirror standard practices in the fast‑moving AI sector. The legal move comes as Apple rolls out public betas of its new Siri AI, intensifying the rivalry between the two tech giants.Leggi di più

OpenAI announced that parents linked to their teen's ChatGPT account will receive notifications if the account is deactivated for violating the company's policies on violent threats or acts. The move expands the parental‑control suite launched last year, which already lets guardians set usage limits, filter sensitive content, and flag self‑harm concerns. Working with online‑safety firm Moonshot, OpenAI also rolled out a new Study Mode and more frequent break reminders for teens who spend long periods in the chatbot. The changes come after criticism that the firm failed to alert authorities when a mass‑shooter’s account was shut down.Leggi di più