Gemini 4 Carbon Tops Leaderboards, Initially Limited to Cyber DefendersNew Decoy Benchmark Reveals Frontier Models Can See But Not ReasonRefactorProof Tests Whether LLM-Generated Proofs Survive Code RefactoringCohortFlow: Steerable Clinical Cohort Curation Agent with Localized Semantic CorrectionStruPO: Faithful Text-to-Image Prompt Editing Without Changing the SceneKnowledge Conflicts Need a Prior, Not a Fixed RuleOpenAI Updates Model Spec with Five Core Values to Guide AI BehaviorOpenAI Details Layered Monitoring to Catch Frontier Model MisalignmentRopes & Gray Builds AI for Transactional Due Diligence with OpenAIStrix Pentest Tool Adds Vercel AI Gateway, Accessing 400+ Models via One APIVercel Flags SDK Starts 7x Faster, Stops Extending Function Durationfx Adapter for AI SDK Runs Harness on Host, Simplifies Sandbox SetupVercel IP Geolocation Now City-Level for Nearly All Requests, Postal Code for 90%YC Runs AI Agent Swarms in Minecraft and MindustrySimantic (YC F26) Lets AI Agents Test Firmware Without HardwareZombie Statistics: Tracing the Origins of Numbers Cited by AIProton Goes All In on AI IntegrationMIT Professor Says AI Is Taking Over Your BrainSkill dreaming: Evolving repository skills through pull request streamsWhen retrieval falls short: stale facts in agent memoryGemini 4 Carbon Tops Leaderboards, Initially Limited to Cyber Defenders

A new decoy counting benchmark exposes a gap between perception and reasoning in frontier models: they can see image content but fail at logical judgments that require ignoring distractors. Elorian’s Glimpse attempts to bridge this by “thinking with images”—generating and manipulating visual representations during reasoning rather than relying on text descriptions alone. The benchmark adds a new dimension for evaluating whether multimodal models truly reason about what they see.

Benchmarks for verifiable code generation typically check a proof against one fixed implementation. A new paper introduces RefactorProof, asking whether the same proof remains valid after behavior-preserving program refactoring. The benchmark derives 458 Lean-certified samples to stress-test the robustness of LLM-generated proofs. The findings point to a practical problem: if generated proofs can’t be reused after refactoring, the engineering value of formal verification drops sharply.

Researchers introduced CohortFlow, a steerable language model agent for clinical cohort curation that represents natural language corrections as localized semantic edits. Instead of reprocessing unchanged eligibility criteria each time requirements evolve, the agent updates only the modified parts. The work was posted as arXiv preprint 2610.36100v1 under the cs.AI new-announcement category. For clinical data teams, this approach could cut the cost of full recomputation when criteria are tweaked, making cohort construction better suited to the iterative nature of real-world research.

An arXiv paper proposes StruPO for training text-to-image prompt editors. The problem: prompt optimization should improve the resulting image without changing the scene the user requested, but stylistic rewriting often weakens object attributes and relations even when global text-image similarity remains high. StruPO uses structural feedback to constrain rewriting and preserve object and relation information from the original prompt. For text-to-image users, this means automatic prompt optimization won’t drift the generated image away from the originally described scene.

An arXiv paper notes that retrieval-augmented language models draw on both retrieved context and their own knowledge, and the two disagree when context is incomplete or wrong. Existing methods settle the disagreement with a rule fixed at design time, such as always trusting retrieved content. The paper argues this is insufficient and advocates using a prior rather than a fixed rule. This means RAG systems handling conflicting information need to dynamically weigh the credibility of both sources based on context, rather than applying a one-size-fits-all rule.

OpenAI published an updated Model Spec outlining five core values meant to guide more capable models toward sound judgment, respect for human agency, and faithfulness to user intent. The spec is OpenAI’s public statement of how its models should behave, covering refusals, safety boundaries, and decision priorities. For developers, it sets expectations for the behavioral boundaries they can rely on when calling OpenAI models.

OpenAI disclosed how it detects frontier model misalignment: a layered monitoring system designed to let responders intervene before incidents escalate when models show unauthorized agent behavior or deviate from intended goals. The approach covers multiple levels of model behavior rather than relying on a single detection method. For developers building agent applications on OpenAI models, this monitoring logic defines when the platform may step in or restrict model behavior.

International law firm Ropes & Gray is developing AI for transactional due diligence with OpenAI. Lawyers review grouped documents, summaries, and reports produced by the AI before advising clients. The tool’s core function is automatically categorizing and summarizing due diligence documents, turning materials that previously required manual review into structured output. Lawyers retain final review, while AI handles initial screening and organization. For the legal industry, this means the document-processing phase of due diligence in large transactions can be compressed.

Strix, the open-source AI penetration testing tool, now runs on any of the 400+ models available through Vercel AI Gateway. It reaches models from Anthropic, OpenAI, Kimi, GLM, and Deepseek through a single API with no markup, removing the need to sign up with multiple providers separately. Usage and costs are consolidated in one place. For security testers, switching models no longer requires new keys or separate billing setup.

Vercel announced that its Flags SDK now initializes 7x faster on cold starts, and usage reporting no longer keeps functions running after the response is sent. The speedup comes from two changes: the SDK loads and parses the flags datafile bundled with the deployment directly, eliminating extra network requests. Previously, usage reporting continued running functions after responses were sent; that behavior is now removed. For developers, feature flags via Vercel Flags have lower latency, and billed function duration is no longer extended by background reporting tasks.

The fx adapter for the AI SDK’s HarnessAgent now uses libfx, the TypeScript library for fx. The harness and model requests run in the application’s process, while built-in filesystem and shell operations remain isolated in the sandbox. The direct benefit is simplified sandbox setup: no additional components need to be installed to run. For developers using HarnessAgent, the debugging path is shorter since harness logic runs in the same process as their application, while dangerous operations stay confined to the sandbox.

Vercel upgraded its IP geolocation to resolve visitor locations at finer network granularity: city-level data for nearly all requests and postal-code data for 90% of traffic. Geolocation data now refreshes more often, keeping locations accurate as IP ranges are reassigned and routing changes. For developers, this means localization, regional compliance checks, and fraud detection can rely directly on platform-provided location data, cutting the maintenance cost of self-hosted IP databases.

Y Combinator Labs published a project running AI agent swarms in Minecraft and Mindustry. The Hacker News post drew 115 points and 43 comments. Game environments offer test scenarios for multi-agent collaboration, resource management, and real-time decision-making that are closer to real-world complexity than single-task benchmarks. Specific experiment designs and results weren’t detailed in the post summary.

The Simantic team (Ahnaf and Seungmin) launched on Hacker News with a simulation layer that lets engineers and AI agents test and validate firmware without physical hardware. Both founders are former firmware engineers at Neuralink and Tesla. The product is a firmware simulation environment where agents can run tests and validate behavior without occupying real devices. For firmware teams, this means testing can be parallelized without hardware queues, and AI agents can participate directly in firmware validation.

This article explores how statistics cited by AI in its responses often have unclear origins and lack reliable sources—a problem the author calls “zombie statistics.” Using Trakkr research, the article traces where the numbers cited by AI actually come from, revealing how these data become detached from their original sources as they spread, are repeatedly cited without verification, and ultimately form “zombie” data that appears authoritative but is in fact distorted. The article provides specific traceability analysis showing that statistics cited by AI often point to second- or even third-hand sources, with original provenance difficult to verify. This research helps readers understand the credibility risks of data in AI-generated content and reminds them to carefully verify statistical information provided by AI. The original link is https://trakkr.ai/trakkr-research/zombie-stats, which received 138 points and 60 comments on Hacker News.

Proton, known for privacy-focused services, is going all in on AI. A Hacker News discussion drew 42 points and 15 comments, with debate over how a privacy product can add AI while keeping its data-protection commitments. Specific feature details weren’t laid out in the post, but the “all in” framing suggests core products like mail and calendar are affected. The key question for Proton users is whether AI features change its end-to-end encryption privacy model.

The Boston Globe published an opinion piece by MIT professor Eric So on AI’s impact on cognitive ability. The core argument is that using AI tools is changing how people think and may weaken independent reasoning and memory. The piece is commentary, with the author raising concerns based on his research field. For general users, the question it raises is worth attention: what does routinely outsourcing thinking tasks to AI mean for cognitive ability in the long run.

arXiv:2610.36188v1 announces a new paper. Large language model (LLM) agents increasingly rely on reusable skills for complex software engineering tasks, making skill libraries critical to repository development and maintenance. However, repository-specific skills must evolve alongside changing codebases. The study proposes “skill dreaming,” a method that evolves repository skills through pull request streams, enabling skill libraries to continuously adapt to repository evolution and thereby improving LLM agent performance on repository development and maintenance tasks.

arXiv:2610.36308v1, announcement type: new paper. Abstract: Benchmarks for memory-based agents typically merge retrieval and answer generation into a single score, making it hard to tell whether a failure stems from missing an update or from failing to use a retrieved update. We examine this distinction on sing.