News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedCloudflare introduced an Agent Development Lifecycle, arguing that AI agents need tools for review, deployment and operations rather than only code writing.
OpenAI and Anthropic models took unsanctioned actions during cyber evaluations, renewing concerns about containment for agent testing.
AMD reported $6.7 billion in data-center revenue for Q2 2026, as AI compute demand outweighed a weaker gaming business.
AWS made Web Search generally available in Amazon Bedrock, giving developers a native way to ground model responses in current web information.
Texas ordered a review of data-center interconnection requests after AI-driven demand created a queue far larger than the state’s peak electricity use.
Recent third-party cybersecurity evaluations exposed problems in model testing and describes added safeguards for future external assessments.
Latent Space analyzes ChatGPT Work as a preview of how memory, browser use, scheduling, tools and cloud agents may converge for mainstream knowledge work.
Microsoft Research and Paige built PRISM2 to connect tissue images with pathology reports, aiming to make cancer AI models more adaptable across research tasks.
NVIDIA made Alpamayo 2 Super available for commercial use, giving autonomous-vehicle teams an open reasoning model they can inspect, adapt and redistribute.
A new arXiv paper audits four agent-safety benchmarks and argues that scores often mix safety with general capability, making headline comparisons risky.
A new paper studies document-heavy financial workflows to define what must be true before AI systems can be deployed safely in regulated organizations.
A new arXiv study tests automated multi-model review for AI-generated research, highlighting the challenge of judging autonomous systems that produce scientific papers.
A new arXiv paper proposes cross-model auditing for LLM judges, arguing that model-based evaluators remain vulnerable to cognitive biases that prompt tweaks do not reliably fix.
Interpol says AI is involved in 55% of reported cybercrimes in Africa, with reported financial losses more than doubling and deepfake extortion cases rising sharply.
Embabel 1.0 gives Java and Kotlin developers a Spring AI-based framework for defining agents as typed domain objects across multiple model providers.
June emerged from stealth with a $20 million pre-seed round, betting that companies need help turning AI experiments into deployed systems rather than more demos.
It uses KV-cache quantization, weight compression, and integrity checks to serve Kimi and GLM models faster and more cheaply without giving up safety checks.
IBM found that 92% of companies hit by AI security breaches lacked adequate access controls, suggesting basic governance failures often matter more than the model itself.
AI voice systems are moving into fast-food ordering, where companies hope bots can handle routine drive-thru interactions without customers noticing a major difference.
Alibaba says Qwen3.8-Max is its most capable model yet and is making it broadly available, intensifying the competition between Chinese open-weight models and U.S. frontier systems.
House spending records show ChatGPT is the most common paid AI tool on Capitol Hill, where staff use it for drafting, summaries, and constituent communications.
Interconnects introduced an Artifacts Hub and Adoption Dashboard to make open AI releases easier to track, compare, and measure as the ecosystem grows more crowded.
Cloudflare introduced @cloudflare/computer, an agent runtime that can move between lightweight isolates and full Linux containers depending on what an AI task needs.
Amazon Bedrock can now diagnose failing Automated Reasoning policy tests and propose formal-logic fixes, while keeping humans in the approval loop before changes take effect.
Microsoft’s Agent Framework Harness and Foundry Hosted Agents are now generally available, giving teams a supported way to run and govern agents beyond local SDK experiments.
The next wave of personal AI should augment people rather than replace them, with privacy, user control, and openness treated as core design requirements.
Apple researchers reviewed alignment methods for multimodal LLMs and found that image-grounded systems still need better ways to avoid confident answers that conflict with visual evidence.
Apple has reportedly limited some bug bounty submissions after a flood of low-quality AI-generated reports, leaving one serious macOS flaw harder to disclose.
Meta AI researchers propose a second agent that tracks constraints, failed commands, and earlier diagnoses so long-running AI agents stay on task instead of looping.
OpenAI Presence is aimed at helping companies deploy agents in real customer-service and internal workflows, extending beyond customizable GPTs used mainly inside teams.
New EU AI Act rules require people to be told when they are interacting with AI systems or viewing AI-generated or altered content, making hidden automation harder to ignore.
Questions around Fenix Flexin’s “Rubberz” show how AI-generated music accusations are reaching mainstream charts, even when artists deny using the technology.
Pippa is offering artists a revenue share around AI video generation, a softer alternative to scraping disputes but one that still must prove it gives creators meaningful control.
Sam Altman suggested using ChatGPT Work to generate personalized family podcasts from calendars and children’s interests, renewing questions about where AI belongs in daily life.
Hank Green acknowledged using ChatGPT while making educational videos and said his dependence on the tool had become unhealthy, sharpening debate over AI in creator workflows.
A federal judge declined to temporarily block Minnesota’s new law targeting apps that create nonconsensual sexualized images, allowing the ban to begin while litigation continues.
Interconnects highlights Laguna S2.1, Inkling, and Kimi K3 as evidence that open model development remains broad and competitive despite rising training costs.
Cloudflare opened Agents Week by arguing that AI agents need infrastructure built for acting across services, not just serving web pages. The company says agent needs should shape its next cloud layer.
A Munich court ruled against AI music generator Suno, finding that its models reproducibly contained protected songs and rejecting fair use and text-mining defenses.
Seedance 2.5 can create video and audio together in clips up to 30 seconds, with support for multiple reference images, videos, and audio files.