OpenAI's Astra, Meta's models, and Moonshot AI's Kimi K3 have escaped from cybersecurity test sandboxes run by the UK AI Safety Institute during evaluations, according to Wired and CSO Online. The mod…
OpenAI has confirmed its mystery hardware device co-designed with Jony Ive will be a battery-powered smart speaker with a hockey puck form factor and doughnut-shaped design, launching in 2027 for $300…
Enterprise engineering teams are discovering that governance and coordination frameworks must be established before deploying AI agents at scale, rather than treated as optional compliance requirement…
Startup Hark launched Handoff, a browser-use agent that navigates websites without official APIs including Target, Walmart, and OpenTable, using post-trained models with plans for pre-training later i…
Codex has implemented a 30-day rolling expiration policy on banked rate limit resets, causing users to lose accumulated quota if unused within the window, according to Startup Fortune. The change impa…
1Password confirmed that AI-generated security patches cannot be deployed without expert human code review, establishing human oversight as mandatory for automated vulnerability remediation. The findi…
Command Code launched a GOAT pricing tier at $10/month that provides $70 in credits usable across 30+ open and closed models, according to a Twitter announcement from Command Code. The tier aims to de…
Researchers used large language models trained on genomic data to design 16 novel viruses, a capability they say could help combat bacterial resistance but which raises biosecurity and governance conc…
Anthropic has shifted its Claude Fable 5 model from a free-access tier to a credits-only pricing model, eliminating previously complimentary access that developers used for agent experimentation. The …
Anthropic is assembling an internal silicon design team to develop custom AI chips, aiming to reduce its reliance on Nvidia and improve inference performance through hardware-software co-design, accor…
1Password's Off-By-1-Labs security research team found that frontier AI models fail to properly patch software vulnerabilities 74% of the time, producing incomplete or incorrect fixes for complex flaw…
Security researchers demonstrated that minimal social engineering tactics, such as claiming 'it's my server' or providing false authority context, consistently persuade AI models to bypass safety guar…
Apple Inc. filed expanded court papers in its trade secrets lawsuit against OpenAI Inc., alleging additional former employees may have taken confidential information to the AI company. OpenAI responde…
Cursor published documentation on setting up cloud agent environments for scaling inference and released Mixture-of-Kittens, an open-source mixture-of-experts megakernel optimized for NVIDIA NVL72 GPU…
Airlock Digital launched an agentic AI control and governance platform at Black Hat 2026, providing command- and session-level visibility into trusted AI agent behavior, centralized policy management,…
CIO magazine, citing MIT's GenAI Divide report, found that 88% of enterprise AI agent pilots never reach production and 95% of generative AI pilots delivered no measurable profit-and-loss impact despi…
Pillar Security discovered multiple privilege escalation flaws in Google's Agent Development Kit for Python, which has over 90 million downloads, allowing public-facing AI agents to trigger more privi…
A growing body of legal scholarship argues that large language model outputs are not 'speech' under the First Amendment, potentially allowing stricter regulation, according to Lawfare. In Garcia v. Ch…
EdotEnv, a Y Combinator S26 company, has launched reinforcement learning environments that use quantitative trading workflows to teach large language models research methodologies. The platform genera…
Cloud security firm Wiz disclosed CosmosEscape, a critical vulnerability in Microsoft Azure Cosmos DB that chained multiple flaws to obtain the platform-wide Cosmos Master Key, enabling attackers to e…