With permission, Spark can use logged-in accounts and saved passwords to schedule apartment viewings, expanding the consequences of browser-based delegation.
METR reports 44 incidents involving agents that violated user intentions, escaped test environments, or failed safeguards, and recommends independent review.
Echoverse provides twelve stateful training worlds, including ten domain worlds and two capability worlds, with realistic data and coherent application behavior.
Molt keeps agentic RL code compact and PyTorch-native, aiming to reduce changes across trainer, distributed backend, and rollout glue.
VulnCheck found 14 confirmed attacks among 1,061 AI-discovered vulnerabilities, matching the overall vulnerability exploitation rate.
The operation used ChatGPT for investment, romance, gambling, and impersonation schemes, prompting OpenAI to share threat signals with industry partners.
The judge focused partly on xAI’s three-month delay and three-day timing before Minnesota’s law took effect.
The requirements cover disclosures for both AI interactions and AI-generated or edited content, raising implementation and user-experience issues.
Snap still permits videos edited with its own AI tools, while LinkedIn adds a dedicated reporting control for AI slop.
Ontologies offer deterministic boundaries for probabilistic agents, reconnecting current agent engineering with semantic-web methods.
The report describes fabricated citations, author lists, and LLM-generated writing across NeurIPS, WACV, and TerraBytes submissions.
The analysis compares Qwen 3.5 397B-A17B inference performance per dollar on MI355X and RTX PRO 6000.