AI
AI News Flash: Mistral, MiniMax, OpenAI Voice, Lawsuit & Breach
Mistral enters robotics, MiniMax plans 2.7T model, OpenAI launches GPT-Live, plus a landmark lawsuit over school shooting.
Mistral Unveils Robostral Navigate, an 8B Model for Robot Navigation Using One Camera
Mistral has unveiled Robostral Navigate, its first AI model for robot navigation. The 8-billion-parameter model uses only a single RGB camera to guide robots through complex environments, achieving up to a 79.4 percent success rate on the R2R-CE benchmark, outperforming both the best single-camera methods and systems using depth sensors or multiple cameras. The model was built entirely in-house and trained solely in simulated environments using about 400,000 recorded paths across 6,000 virtual spaces, and it works on wheeled, legged, and flying robots. Mistral has not yet announced availability but sees navigation as foundational for universal robotics and plans further improvements. Experiments with reinforcement learning already boosted the success rate by 3.2 percentage points, and the company states that more training and experiments will continue to push performance higher.
MiniMax Plans Open-Source Release of 2.7 Trillion Parameter Model M3 Pro
Chinese AI developer MiniMax is working on a new large language model with 2.7 trillion parameters, internally called M3 Pro, which it plans to release as open source possibly as early as Q3. The model would be significantly larger than its current top model M3, which has 428 billion parameters. Larger models tend to perform better on complex reasoning and multi-step instructions. Chinese open-source models have gained traction with developers this year, especially for high-volume, less critical tasks. MiniMax competes with Zhipu, DeepSeek, and Moonshot AI. Recent reports suggest the Chinese government wants to tighten controls on future releases of such models.
Chinese AI startup MiniMax plans to open-source a 2.7 trillion parameter model later this year →
British Columbia Prepares Lawsuit Against OpenAI Over Failure to Report Violent ChatGPT Activity Before School Shooting
British Columbia announced on Tuesday that it is preparing a lawsuit against OpenAI for failing to report violent ChatGPT activity by the person who committed a mass school shooting in Tumbler Ridge in February 2025. The perpetrator, Jesse Van Rootselaar, had her account banned by OpenAI in June 2025 but the company did not alert law enforcement. Provincial Attorney General Niki Sharma stated the province wants to hold OpenAI accountable and ensure that British Columbians are not left bearing the costs of corporate wrongdoings. Canadian families have already filed lawsuits in a California court, and the province is coordinating with them. OpenAI responded that it has strengthened safeguards and has a zero-tolerance policy for using its tools for violence. CEO Sam Altman apologized in April, saying he was deeply sorry for not alerting law enforcement. Lawyers for the families allege that OpenAI chose to stay silent because reporting one case would mean reporting thousands, and that the company instructed banned users on how to resume usage. Mark Daley, chief AI officer at Western University, noted that no single company should hold such knowledge alone.
Canadian province prepares lawsuit against OpenAI after school mass shooting →
Lone Attacker Used AI Workflows to Breach AWS Environment in 72 Hours
A lone financially motivated attacker used agentic AI workflows to breach a large Amazon Web Services environment in approximately 72 hours, according to incident response firm Sygnia. The attacker chained weaknesses across application services, AWS resources, source code repositories, CI/CD pipelines, runtime components, and data stores, accelerating reconnaissance, tool development, and adaptation. The attacker first obtained an AWS access key through an internet-facing application weakness and then ran systematic secrets theft, backdoor creation, and data exfiltration workflows each time new access was gained. To demonstrate capability, the attacker performed mostly reversible impact actions such as denying S3 bucket access, limiting ECS services, blocking network access, and purging SQS queues. Sygnia’s vice president Avi Dayan emphasized that operationally, mean time to detect and remediate must contract significantly, and security operations should pivot to automated, high-fidelity response playbooks and AI-driven defense. Sygnia recommended comprehensive visibility, strengthened identity security, layered defense controls, and automated detection and response processes, as well as predefined containment procedures.
Lone Attacker Uses AI to Breach AWS Cloud Environment in 72 Hours →
OpenAI Launches GPT-Live with Full-Duplex Voice for More Natural Conversations
OpenAI launched GPT-Live on Wednesday, a pair of new voice models replacing Advanced Voice Mode with a full-duplex architecture that enables simultaneous listening and speaking. GPT-Live-1 becomes the default voice model for paid ChatGPT users on Go, Plus, and Pro tiers, while GPT-Live-1 mini serves free users, rolling out globally on iOS, Android, and ChatGPT.com. The model can make interaction decisions many times per second whether to speak, listen, pause, interrupt, or invoke a tool, and delegates complex reasoning to GPT-5.5 asynchronously, allowing continued conversation. Over 150 million people talk to ChatGPT using voice each week. The update introduces richer safety evaluations: GPT-Live-1 showed improvements over Advanced Voice Mode on illicit behavior (0.63 to 0.97), self-harm (0.72 to 0.98), and hate speech (0.87 to 1.00), with a slight non-significant regression on emotional reliance (0.88 to 0.82). OpenAI built real-time safeguards to steer toward safer responses and surface crisis resources. The company acknowledged limitations: no voice with video or screen sharing at launch, potential non-native accents for certain languages, and no API access on day one.
OpenAI launches GPT-Live, a full-duplex voice upgrade that lets ChatGPT talk more like a person →