OpenAI Pipeline Finds and Patches 948 Linux Kernel Privilege-Escalation FlawsDify 1.19.0 Adds Nested Workflow Tracing Layer by LayerOpenAI Acquires Pierre Computer to Expand Codex Code ToolsOpenAI Opens GPT-5.6 Cyber to AWS GovCloud Customers for Cybersecurity WorkKimi K3.1 Pro Gains More Zero Data Retention ProvidersQualcomm Cloud AI 100 Ultra Packs 128GB at 548GB/s in a 150W CardVoters Reject AI Data Centers Across 45 Ballot Measures in US MidtermsYC's Garry Tan Spends Up to $200K a Year on AI Agent TokensRSI-Dojo Trains LLM Agents to Automate LLM Post-TrainingEvoGroup: A Self-Evolving Harness for Multi-Agent Teams Sharing ContextEl Agente Fármaco: An AI Scientist That Chains Drug Discovery ToolsLOLPlay Benchmark Tests Real-Time Visual Agents Across 36 League of Legends Champion TasksDory1.0-4B: A 4B Looped Model That Adjusts Depth to Compute BudgetFlagRot: Selective Rotation for Low-Bit KV CachesKino AI Bundles Media Management and Editing So Claude Code Can Edit TimelinesWorkflow SDK Adds Atomic Starts to Eliminate Duplicate-Run RaceOpenClaw Ships 2026.11.5-beta.1, Fixes OpenRouter Billing Failuredatasette-llm 0.1a10 Fixes Model List Ignoring ConfigOpenAI Codex Ships 0.168.0-alpha.2 After Multiple Alpha BuildsOpenAI Pipeline Finds and Patches 948 Linux Kernel Privilege-Escalation Flaws

Dify released 1.19.0, bringing nested workflow tracing into observability tools. Users can follow a workflow into the Workflow Tools it calls, inspect steps inside each invocation, and keep LLM context alongside the execution tree. The release also makes Workflow Tools portable with App packages and adds Skill H capabilities. Teams building multi-level agent flows can now pinpoint which sub-workflow step failed instead of seeing only a top-level error.

OpenAI will acquire Pierre Computer Company, Inc., integrating its tools into Codex for building, understanding, and reviewing code. OpenAI did not disclose the deal value, team size, or an integration timeline. For Codex users, future releases may gain stronger code review and code comprehension capabilities.

OpenAI has made Daybreak Red available in AWS GovCloud (US-East), delivering its GPT-5.6 Cyber model to approved customers through Amazon Bedrock for authorized cybersecurity work. The deployment targets government cloud environments, letting federal and regulated agencies run security analysis and vulnerability triage on a dedicated model inside compliance boundaries rather than through public API channels.

Kimi K3.1 Pro and its faster variant, Kimi K3.1 Pro Highspeed, are now available on AI Gateway from more Zero Data Retention (ZDR) providers. Until now, Moonshot AI was the only provider serving either model. K3.1 Pro improves on earlier Kimi models in coding, 3D modeling, and financial analysis. For teams with data compliance requirements, multiple ZDR providers mean they can use Kimi models without data being retained by third parties, while gaining failover options.

Qualcomm has published specs for its Cloud AI 100 Ultra data center accelerator: 150W per card, 128GB of memory, and 548GB/s of bandwidth. The combination puts large memory capacity into a low-power envelope suited to inference deployments in power-constrained racks. For data centers running model inference under tight power and cooling budgets, it offers an alternative to Nvidia hardware.

Tom’s Hardware reports that 45 ballot measures concerning AI data centers went to voters in the US midterm elections, and Americans overwhelmingly voted against the developments, with data centers losing across the board. The outcome directly affects project timelines, as siting and permitting now face stronger local resistance, forcing operators to rethink community engagement and power infrastructure plans.

TechCrunch reports that Y Combinator’s Garry Tan uses AI agents so heavily that his token spend reaches up to $200,000 a year. The figure makes the cost of heavy agent use concrete: a single account’s annual outlay is on the order of an employee’s salary. For companies planning to roll out agents at scale, it is a real-world cost reference, and it shows agent call volume among top users has reached substantial levels.

An arXiv paper introduces RSI-Dojo, aimed at training agents to automate LLM post-training. Recursive Self-Improvement (RSI) refers to AI systems autonomously building and improving models, harnesses, and underlying infrastructure; the paper targets a central capability—agents that perform training tasks. Only the abstract is available, with no benchmark results, model scale, or training cost disclosed. If the method works, part of the manual effort in post-training could shift to agents.

A new paper introduces EvoGroup, aimed at group-agent setups such as Claude Tag where agents act as digital workers in long-running team collaboration. Their shared contexts accumulate continuously through multi-user interaction, which existing approaches handle poorly. EvoGroup uses a self-evolving harness to manage group agents within shared contexts, letting multiple agents collaborate in one workspace over time without stepping on each other.

Computational drug discovery offers a growing range of tools for proposing and evaluating molecules, but combining them and deciding what to do next still demands substantial expert effort. The paper introduces El Agente Fármaco, a scalable AI scientist that automates tool orchestration and decision-making. Its value is freeing experts from workflow scheduling so molecule screening iterations run faster and cover more candidates.

Researchers have released LOLPlay, a benchmark for evaluating visual agents in the native League of Legends environment. It includes 36 champion-specific tasks spanning resource management, combat, and cooperation. Unlike static tests, the environment evolves during inference, requiring agents to coordinate reasoning and action in real time. The work is published on arXiv as 2611.03685. For developers, LOLPlay offers an evaluation standard grounded in a genuinely dynamic setting, useful for exposing weaknesses in long-horizon decision-making and real-time reaction of visual agents.

Researchers released Dory1.0-4B, a 4-billion-parameter looped (latent-recurrent) language model with elastic depth, so a single checkpoint serves different inference compute budgets. It is built on a normalized Transformer (nGPT) backbone with an interleaved layer structure. The practical payoff is deployment-side: one set of weights can shift between low- and high-compute settings depending on the scenario.

Long-context inference is bound by the KV cache, and 2-bit codes could shrink it roughly sixfold if the error stays in check. The standard remedy rotates keys before quantizing on the assumption that rotation always helps, but the paper observes that it does not. FlagRot applies rotation selectively, only where it actually reduces error, keeping quantization error contained at low bit widths.

Kino, a YC S23 company, launched on Hacker News as an AI media manager and video editor. It bundles footage and asset storage, a full editor, and review links for teammates into one product. The distinguishing feature is agent participation in editing: Codex and Claude Code can work on the same timeline alongside colleagues. For teams mixing people and multiple AI agents on one video, it removes the step of shuttling assets between separate tools.

Vercel’s Workflow SDK now makes start() an atomic operation, admitting a run and claiming a hook in a single step, so duplicate workflow starts fail safely. The change replaces the old getHookByToken()-then-start() pattern, which had a documented duplicate-start race between lookup and start. Developers get predictable behavior under retries and concurrency, and can drop a layer of manual deduplication at workflow entry points.

OpenClaw released 2026.11.5-beta.1 alongside several release-publish builds. Changes include a fix for a 2026.11.1 OpenRouter billing failure (#192936), a scoped Kova waiver for 2026.11.1 (#192812), and a UI fix localizing skill and plugin feedback (#192305). The notes also mention reading waiver aggregate logs with escape sequences allowed. Teams using OpenClaw with OpenRouter should see the billing failure resolved after upgrading.

Simon Willison released datasette-llm 0.1a10, fixing a bug that affected the agent.datasette.io demo site, where the list of available LLM models was not correctly restricted to those in the configuration. The release is an LLM plugin for datasette. Developers building agent demos with the plugin will now see model visibility match their config, avoiding exposure of unauthorized models.

OpenAI Codex released 0.168.0-alpha.2, following a run of updates including 0.167.0-alpha.3.1, 0.167.0-alpha.5, and 0.168.0-alpha.1. The release note lists only the versions, with no specific feature changes or fixes disclosed. For Codex users, this is a routine alpha-channel iteration; stable-release details remain unannounced.