OpenAI's Patch the Planet Uses GPT-5.6 Sol to Find PHP2Shell Flaws Affecting 70% of the WebAnthropic Donates Agent Skills Standard to Linux Foundation's AAIFPaper Examines AI and the Enclosure of the Scientific CommonsAI CT Analysis Finds Coronary Inflammation in 42.9% of PatientsNew Paper Studies KV Cache Retention Policies in Disaggregated LLM ServingReAPO Cuts Repeated Failed Strategies in LLM Reasoning TrainingSimHack Predicts Reward Hacking Before RL Training StartsCan Frontier Coding Agents Do Real Science? Study Finds GapsAI Research Teams Autonomously Search for Linear Attention ArchitecturesDiffResearch Uses Diffusion Language Models to Cut Deep Research LatencyDeepMind Maps Global Forest Types at 10-Meter ResolutionReplit Launches Monitoring Agent to Watch Apps for Solo DevelopersOpenAI Launches Decisions API for Fast, Low-Cost App DecisionsModal CEO Erik Bernhardsson on Why AI Compute Comes in RectanglesVercel Blog Argues for Ending Long-Lived Static CredentialsArmin Ronacher: 30 Years of Bad Software Is Now Being Punished by AI Security BugsJames Yu Curates Award-Winning AI Literature Reading ListTypeUI Releases UI Micro-Interaction Prompt Collection for AIPoll: Most Americans Say AI Hurts Society, Oppose New Data CentersVercel Adds 8 New TLDs to Free Domain Offer for Pro TeamsOpenAI's Patch the Planet Uses GPT-5.6 Sol to Find PHP2Shell Flaws Affecting 70% of the Web

Anthropic announced it is donating the Agent Skills standard to the Agentic AI Foundation (AAIF) under the Linux Foundation, which Anthropic co-founded with Block and OpenAI. The company said the move keeps Agent Skills open, community-driven and vendor-neutral. Developers can now build Agent skills against a single standard that works across vendors.

An arXiv paper frames AI’s impact on scientific knowledge production and sharing as an ‘enclosure of the scientific commons.’ It drew 104 points and 59 comments on Hacker News, the highest engagement in this batch. Discussion centered on training data provenance, open access to papers, and model vendors’ control over research infrastructure. For readers tracking the relationship between AI and the academic ecosystem, it is an argument worth following.

Biograph released results from the first U.S. study of Caristo Diagnostics’ AI-powered coronary inflammation detection, finding the condition in over 42% of asymptomatic patients—42.9% specifically. The result indicates AI can surface cardiovascular risk signals in people without obvious symptoms. If validated further, such detection could enter routine screening imaging workflows, turning previously overlooked CT data into risk alerts.

A new arXiv paper examines KV cache retention in disaggregated LLM serving, where prefill and decode run on separate node pools and a network fabric sits between cache computation and consumption. This breaks the classical cache assumption that a miss only incurs local recomputation cost. The work studies when to keep versus evict KV cache to balance network transfer against recomputation. The findings are directly relevant to teams running disaggregated inference, affecting latency and bandwidth costs.

ReAPO (Retry-Aware Policy Optimization) targets a known inefficiency in RLVR training for LLM reasoning: independent stochastic rollouts often resample similar failed strategies. The method makes exploration aware of prior retries, so sampling budget is not wasted on strategies that already failed, improving exploration efficiency. This is a training-side optimization for reinforcement learning, directly affecting the cost and convergence speed of reasoning model training. It is relevant to teams running their own RL pipelines.

The paper introduces SimHack, which predicts the onset of detectable reward hacking before reinforcement learning training begins. The authors cite two practical difficulties: environment hardening gets harder as training environments grow more complex, and training-time interventions are limited in what they can fix. Moving the problem to pre-training prediction lets teams assess reward design risk before spending compute. For teams training code models with RL, it offers a low-cost early screening step.

A new paper tests the claim that frontier coding agents can act as end-to-end automated researchers, using ML-for-science tasks as a demanding benchmark. The finding: such agents remain limited by the need for domain insight that goes beyond routine ML engineering. Improving a scientific model often requires knowledge outside standard coding and tuning workflows. The authors introduce an agent framework aimed at closing that gap. For teams treating agents as research accelerators, the result suggests keeping human checkpoints for domain expertise.

The paper explores using teams of collaborative AI researchers to autonomously discover linear attention architectures, with agents dividing work across idea generation, code modification, and experiment execution. To avoid getting stuck in local minima, the authors design mechanisms that encourage diversity in the search. The significance lies in shifting architecture search from manual trial-and-error to a parallelizable automated process. If the approach generalizes, the pace of new architecture discovery may no longer be capped by human headcount.

A new arXiv paper, DiffResearch, notes that deep research systems iteratively plan, gather evidence, synthesize and revise, making end-to-end latency a key bottleneck for interactive use. Nearly all production systems rely on autoregressive language models and inherit left-to-right decoding. DiffResearch instead applies diffusion language models to speed up the pipeline. The abstract does not state a specific speedup figure.

Google DeepMind published a global map that classifies every 10-meter patch of land on Earth into distinct forest types. The work goes beyond conventional tree-cover statistics by categorizing forest types at fine granularity. The map’s resolution and classification detail have direct applications in ecological monitoring, carbon accounting and land-use research. Technical methods and validation data were not disclosed in the summary.

Replit launched a Monitoring Agent aimed at indie developers and small teams who publish apps but have no on-call engineer when users report ‘it doesn’t work for me.’ Large companies page an on-call engineer to investigate; small teams usually have only the builder. The agent monitors app health and surfaces problems. Replit did not disclose which metrics are tracked or how alerts are delivered. For solo developers, it adds an automated on-call helper.

OpenAI released the Decisions API, positioned to provide fast, focused and low-cost decision-making for any application. The company emphasizes response speed and low call cost, targeting scenarios that require frequent judgments. Specific latency figures and pricing details were not disclosed. For developers, it offers a way to offload simple decision logic to an API instead of invoking a full large model each time.

Modal CEO Erik Bernhardsson joined Latent Space to discuss why AI compute clusters take rectangular shapes, how elastic GPUs are scheduled, how sandbox usage is scaling, and the economics of the agent cloud. The conversation spans hardware layout through business models, focusing on how agent workloads demand infrastructure that differs from traditional cloud services. For developers building agent products, it covers cost structure and infrastructure choices.

A Vercel blog post traces how TLS went from protecting only the most sensitive parts of an app 15 years ago to becoming the default for every request, driven by Google, Let’s Encrypt, the EFF and others pushing HTTPS everywhere. The post draws a parallel to long-lived static credentials today, noting that anyone between browser and server can read and tamper with unprotected traffic, and argues for short-lived credentials and automatic rotation as the default.

Simon Willison’s blog quotes Armin Ronacher: “As fun as AI is, we built so much bad software over the last 30 years, and now the universe is punishing us for our sins.” The quote is tagged security, generative AI and AI security research, pointing to legacy software quality problems exposed as AI accelerates vulnerability discovery.

Developer James Yu published an ‘award-winning AI literature’ list on his blog, collecting widely cited and discussed AI papers with reading guidance. The post drew 57 points and 60 comments on Hacker News, with discussion focused on which papers are genuinely worth close reading. For readers looking to build AI fundamentals systematically, the list offers a ready-made entry path that saves time filtering through arXiv.

TypeUI posted a collection of UI micro-interaction prompts for AI at typeui.sh/micro-interactions, aimed at developers generating interface details. The Hacker News submission drew 68 points and 42 comments. It is aimed at front-end developers who need to quickly produce button, loading and transition animations.

An ABC News/Ipsos poll found that most American respondents view AI’s impact on society as negative and oppose building new data centers. The results contrast with the tech industry’s continued heavy investment in AI infrastructure. The survey also covers public attitudes toward data centers, suggesting compute expansion is meeting local resistance. Sample size and margin of error require checking the original report.

Vercel expanded its free one-year domain offer for Pro teams with eight new TLDs: .co, .studio, .company, .info, .pro, .live, .digital and .agency, bringing the total eligible options to 16. Each team can still claim only one free domain, and premium domains are excluded. After the first year, standard renewal pricing applies. The change gives indie developers and small teams more naming flexibility beyond crowded .com options.