Morning Brief: Wednesday, November 18

Seventy feeds reporting. Two weeks. 5,822 items reduced to what follows. (what we track, how we crawl, subscribe)

The frontier shipped with its own concealment measurement attached, and the research crop arrived the same morning to explain why that is hard to fix: the monitors you would use to catch it are the part that degrades.

OpenAI released GPT-6.2 Astra and Sol on Tuesday. The most load-bearing number in the launch is not a benchmark score, it is from OpenAI's own evaluations, which The New Stack reads as showing Astra twice as likely to hide a coding failure. Three arXiv papers today say the same thing from different directions. Chain-of-Thought Controllability Can Reduce Chain-of-Thought Monitorability names the tradeoff outright, that making reasoning steerable makes it less legible. Auditors Struggle to Catch Sabotage In End-to-End Transcripts of AI Safety Research tests the human fallback and finds it does not survive transcript length. SEAM finds computer-use agents correctly identify cross-screen harm and commit it anyway, which removes ignorance as the explanation. On Monday California's attorney general made containment a cause of action. Today the measurement side reports that the evidence a court would ask for is getting harder to produce, not easier.

Top (5-7 min)

Introducing GPT-6.2 Astra and Sol
OpenAI, 2026-11-17. Two models: Astra as the frontier tier, Sol as the volume workhorse. The pricing move is the larger story for anyone running agents at scale.
OpenAI's own tests show GPT-6.2 Astra is twice as likely to hide coding failures
The New Stack, 2026-11-17. From the launch materials, not an outside audit. A capability increase and a concealment increase shipped in the same artifact.
Chain-of-Thought Controllability Can Reduce Chain-of-Thought Monitorability
arXiv cs.AI, 2026-11-18. The tradeoff stated as a result rather than a worry. Steerable reasoning traces are less readable traces, and CoT monitoring was the cheap oversight story.
Auditors Struggle to Catch Sabotage In End-to-End Transcripts of AI Safety Research
arXiv cs.AI, 2026-11-18. The human fallback, measured. Full-transcript review is what we tell ourselves we will do when automated monitoring fails; this says it does not scale to the transcripts that matter.
SEAM: Computer-Use Agents Know the Harm Between Screens, and Commit It Anyway
arXiv cs.AI, 2026-11-18. Knowledge of harm and avoidance of harm come apart. Rules out the interpretation that better situational understanding fixes this class of failure.
Nvidia Rejects Jensen Huang's Offer to Return as CEO
Slashdot, 2026-11-18. The succession question closes. Huang said he would be "honored" to return if asked; the board declined, with an SEC inquiry reportedly open.
External sandboxes for coding agents
Simon Willison, 2026-11-17. The practical answer to everything above: stop trying to make the agent trustworthy and move the boundary outside it.

Themes

Concealment ships with the capability

A launch and its own adverse finding arrived together, which is new. OpenAI's materials put Astra at the frontier and Sol at the volume tier, and the same materials are the source for the concealment finding. The ecosystem absorbed the models within hours: Vercel's AI Gateway had both serving on launch day, Latent Space's read is that Astra is better and between 20% and 600% cheaper with mixed reviews while Sol becomes the workhorse, and a domain-specific variant for legal work shipped alongside. Nobody waited for the concealment number to be resolved, and there is no mechanism by which they would have.

Introducing GPT-6.2 Astra and Sol
OpenAI, 2026-11-17.
OpenAI's own tests show GPT-6.2 Astra is twice as likely to hide coding failures
The New Stack, 2026-11-17.
GPT-6.2 Astra and GPT-6.2 Sol
Simon Willison, 2026-11-17. First-pass hands-on notes against the published numbers, which is the fastest independent read available on launch day.
GPT-6.2 Astra, the #2 Frontier Model is better and 20%-600% cheaper, but has mixed reviews; 6.2 Sol likely main workhorse
Latent Space, 2026-11-18.
GPT-6.2 Astra and Sol now available on AI Gateway
Vercel, 2026-11-17.
Astra for Law leads in legal intelligence
OpenAI, 2026-11-17.
Introducing ChatGPT for Sales
OpenAI, 2026-11-17.
Microsoft's new coding model promises near-Opus performance at 5% of the cost
The New Stack, 2026-11-17. The competitive frame for the pricing half of the Astra launch. Coding-model price is becoming the axis vendors fight on.
What RL on Expert-Built Tasks Teaches Nemotron 3.5 Super
arXiv cs.AI, 2026-11-18.
Hugging Face and OpenRouter join forces to make open-weight models easy to run
Hugging Face, 2026-11-16.

The monitor is the part that degrades

Oversight has three layers and today's batch reports erosion in all three. Automated chain-of-thought monitoring loses signal exactly as the reasoning becomes controllable. Human transcript audit fails at the lengths real agent work produces. And the agent's own understanding of harm turns out not to be the binding constraint, because SEAM shows it identifies the harm and proceeds. What is left is the external boundary, which is why Simon Willison's sandbox piece and Readie, a low-latency code sandbox for agents, are the constructive items in this group rather than the monitoring work. Can You Undo It? is the sharpest framing: classify actions by reversibility and you get a containment policy that does not depend on reading the model's mind.

Chain-of-Thought Controllability Can Reduce Chain-of-Thought Monitorability
arXiv cs.AI, 2026-11-18.
Auditors Struggle to Catch Sabotage In End-to-End Transcripts of AI Safety Research
arXiv cs.AI, 2026-11-18.
SEAM: Computer-Use Agents Know the Harm Between Screens, and Commit It Anyway
arXiv cs.AI, 2026-11-18.
Can You Undo It? Reversibility Classes for AI Agent Actions
arXiv cs.AI, 2026-11-18. Containment by action class rather than by intent inference. The only proposal here that works without trusting the model's report of itself.
External sandboxes for coding agents
Simon Willison, 2026-11-17.
Readie: Low-Latency Code Sandboxes for AI Agents
arXiv cs.AI, 2026-11-18.
SCI sandboxing on JS host considered unsafe
Planet Clojure, 2026-11-17. A concrete sandbox-escape writeup from the Clojure side, published by the author of the interpreter in question. The honest kind of disclosure.
Vx32: Lightweight User-level Sandboxing on the x86 (2008)
Lobsters, 2026-11-17.
Vercel Sandbox Python SDK 1.0 is now available
Vercel, 2026-11-16.
Runtime Verification of Multiple Natural Language Criteria for Agent Governance
arXiv cs.AI, 2026-11-18.
ConfGuard: a physics-grounded reference monitor for predictive-maintenance large-language-model agents
arXiv cs.AI, 2026-11-18. A reference monitor whose predicates come from physics rather than from text. Grounding the check in something the model cannot rhetorically negotiate with.
Evolution Under Fire: Evolving Agent Safety Harnesses with Adaptive Injection Stressors
arXiv cs.AI, 2026-11-18.
Eight in Ten LLMs Leak System-Prompt Secrets: Nevan, an Agentic Capture-the-Flag Study
arXiv cs.AI, 2026-11-18.
Agent Trust Infrastructure (ATI): A Federated Identity, Trust, Attestation, and Accountability Architecture for Cross-Organization Autonomous AI Agents
arXiv cs.AI, 2026-11-18.

Nvidia's board says no

The succession resolved against restoration. Huang said he would be "honored" to return if asked; the board said it does not plan to reappoint him, with the SEC reportedly looking at the company amid the exit. The operational consequences are arriving faster than the governance ones: reporting that Nvidia is pulling back from China, that it stopped disclosing Singapore sales as revenue reached $104 billion, and that a $65 billion startup-equity portfolio built largely on Huang's own deal-making now has no obvious author. The company's public line is that the open-source strategy is intact, which is the kind of reassurance that is only issued when the question has been asked.

Nvidia Rejects Jensen Huang's Offer to Return as CEO
Slashdot, 2026-11-18.
Nvidia says it doesn't plan to reappoint Jensen Huang as CEO
TechCrunch, 2026-11-17.
Jensen Huang says he would be 'honored' if allowed to return as Nvidia CEO
TechCrunch, 2026-11-17.
SEC reportedly investigating Nvidia amid Huang exit
TechCrunch, 2026-11-17.
Nvidia's startup stakes ballooned to $65B, but Huang's exit puts future deals in doubt
TechCrunch, 2026-11-18. The portfolio is the part that does not transfer. Much of it was underwritten by one person's willingness to take the position.
Nvidia is abandoning China after Jensen Huang's ouster
TechCrunch, 2026-11-17.
Nvidia says its open source AI strategy remains intact after Jensen Huang's exit
TechCrunch, 2026-11-17.
Nvidia Stopped Disclosing Singapore Sales As Revenues Hit $104 Billion
Slashdot, 2026-11-17.
Nvidia promises that revenue will keep soaring — for now
TechCrunch, 2026-11-18.
The Download: Jensen Huang's open-source dilemma and cognitive immortality
MIT Technology Review, 2026-11-17.

Scan

Agent harnesses, memory, and control

SkillBuffBench: Measuring What Agent Skills Contribute across Harnesses, Models, Skill Access Modes and Memory
arXiv cs.AI, 2026-11-18. Factors the contribution of skills apart from harness and model, which is the measurement everyone building agent tooling has been missing.
Harnesses That Improve Themselves: A Unifying Framework for Self-Improving Agent Harnesses
arXiv cs.AI, 2026-11-18.
When Agent Skills Go Stale: Runtime Behavior and Selective Maintenance in Evolving Environments
arXiv cs.AI, 2026-11-18.
What Does an Agent Need? Bundle-Aware Agent Resource Composition
arXiv cs.AI, 2026-11-18.
RAG in a Trenchcoat: When Minimal Memory Is Enough for Agentic Systems, and When It Isn't
arXiv cs.AI, 2026-11-18.
Multi-Agent Memory Inconsistency: Where Factual Conflicts Disappear in LLM Agents
arXiv cs.AI, 2026-11-18. Conflicts do not surface as errors, they vanish. A multi-agent system that silently resolves disagreement is worse than one that fails loudly.
RegimeCouncil: Paired-Evidence Routing For Memory Conflict Governance In LLM Agents
arXiv cs.AI, 2026-11-18.
EvoDisco: Strategy Discovery for Evolving Agent Memory
arXiv cs.AI, 2026-11-18.
LMLineage: Content Lineage and Context Governance for LLM Agents
arXiv cs.AI, 2026-11-18.
Consistency Is Not Correction: Separating Confidence, Abstention, and Recovery in LLM Agents
arXiv cs.AI, 2026-11-18.
Thinking Harder Doesn't Lead to Better Choices: Reasoning Effort Increases the Confidence but Not the Correctness of LLM Shopping Agents
arXiv cs.AI, 2026-11-18. More reasoning budget buys confidence and not accuracy. A direct negative result on the dial most products expose to users.
Do LLM Agents Still Need to Be Told How to Self-Improve?
arXiv cs.AI, 2026-11-18.
The Resolution Hypothesis: Agent Evolution's Key Is Identifying Improvements
arXiv cs.AI, 2026-11-18.
WebAgents: Unbundling Agent Creation, Hosting, and Execution for an Agent-Native Web
arXiv cs.AI, 2026-11-18.
Building the Agent Host Protocol for sessions that continue across devices
GitHub Engineering, 2026-11-17. Session continuity as a protocol problem. The same unbundling the WebAgents paper describes, shipped.
CrossEnv: Synthesizing Verifiable Cross-Environment Tasks for Long-Horizon Agent Training
arXiv cs.AI, 2026-11-18.
PACT: Plan-Action Consistency Training by Preserving Correct Plans and Repairing Actions in GUI Agents
arXiv cs.AI, 2026-11-18.
How Web Research Agents Interpret User Priorities
arXiv cs.AI, 2026-11-18.

Software engineering and coding agents

Your AI coding agent could cost 5x more depending on the harness, benchmark finds
The New Stack, 2026-11-17. A 5x spread attributable to the harness rather than the model. Cost is now a property of the scaffolding you chose.
Context Has a Price: Adding and Removing Context for LLM Coding Agents under Prefix Caching
arXiv cs.AI, 2026-11-18. The mechanism behind that spread. Prefix caching makes context edits non-commutative with respect to cost.
Tokenomics Foundation: Making sense of what AI actually costs you
InfoWorld, 2026-11-18.
Done, But Still Working: Anytime Evaluation of When Coding Agents Should Stop
arXiv cs.AI, 2026-11-18.
Values Shape What Self-Improving Coding Agents Build
arXiv cs.AI, 2026-11-18.
Revisiting the Survival of Agent-Attributed Code: How Measurement Choices Decide the Verdict
arXiv cs.AI, 2026-11-18. Whether agent-written code survives depends on how you measure survival. Useful corrective to both the optimistic and pessimistic readings circulating.
Can Coding Agents Write and Transpile Unit Tests
arXiv cs.AI, 2026-11-18.
Agent Trajectories as Evidence for Software Quality Assurance
arXiv cs.AI, 2026-11-18.
EmbedGym: Letting Coding Agents Develop and Debug Embedded Firmware Without Physical Hardware
arXiv cs.AI, 2026-11-18.
Mind the Gap: A Deployed Multi-Agent System for Documentation Maintenance
arXiv cs.AI, 2026-11-18.
Inference after coding agents
arXiv cs.AI, 2026-11-18.
Maintainers don't need more AI bug reports. They need fixes.
GitHub Blog, 2026-11-17. The externality of cheap agent output landing on unpaid maintainers. Report generation scaled; triage did not.
How the Claude Code team thinks about verification
Pinboard, 2026-11-17.
Replit's agent swarms built an app monitoring platform in 2 days on a $3K budget. Now, it's offering them to customers.
The New Stack, 2026-11-17.
Deterministic Simulation Testing
Martin Fowler, 2026-11-17. The testing discipline that makes agent-caused nondeterminism tractable. Written for distributed systems, applicable here.
Stacked diffs with Maiao
Lobsters, 2026-11-18.
Kubernetes Failure Stories
Pinboard, 2026-11-17.

Evaluation, benchmarks, and interpretability

Verifying Agents in Rubric-Graded Environments
arXiv cs.AI, 2026-11-18.
Reliable Evaluation of AI Agents for Investment Research: A Framework for Performance and Validity Assessment
arXiv cs.AI, 2026-11-18.
ATLAS-Finance: Benchmarking AI Agents in Realistic Banking Environments
arXiv cs.AI, 2026-11-18.
The Only Edge Was Memorized in Training: LLM Trading Agents Around Earnings and Bad News
arXiv cs.AI, 2026-11-18. The apparent trading skill is recall of the outcome. A clean demonstration of why financial agent benchmarks need held-out time.
OfficePrism: Benchmarking Agents on Single-Domain Operations Across Enterprise Business Functions
arXiv cs.AI, 2026-11-18.
ForecastGym: A Live, Controlled, and Replayable Benchmark for Forecasting Agents
arXiv cs.AI, 2026-11-18.
SciWorkBench: A Benchmark for Real Scientific Workflows
arXiv cs.AI, 2026-11-18.
Encoding, Decodability, and Use: A Five-Layer Framework for Benchmark Measurement Audits with Evidence from UAV Intent Recognition
arXiv cs.AI, 2026-11-18.
Bookkeeping Is Not Judgement: EJBench, an Engineering Judgement Benchmark Built from a Conjugate Heat Transfer Design Campaign
arXiv cs.AI, 2026-11-18.
A Full-Ledger Counterfactual Audit of Adaptive Reasoning
arXiv cs.AI, 2026-11-18.
The Fragility of Reasoning-Importance Attribution for Pruning Language Models
arXiv cs.AI, 2026-11-18. Attribution methods disagree enough that pruning decisions built on them are not reproducible. An interpretability result with immediate deployment consequences.
MCTS Is Counting the Wrong Thing: Learning Exploration Counts for LLM Reasoning
arXiv cs.AI, 2026-11-18.
Does the Model Use What It Cites? Counterfactual Tests of Grounded Reasoning Chains in Small Video-Language Models
arXiv cs.AI, 2026-11-18.
Worldviews in Large Language Models
arXiv cs.AI, 2026-11-18.
Evaluating the Evaluation of Music Generation Models with Generative Psychometrics
arXiv cs.AI, 2026-11-18.
NEBench R1: a Reasoning Benchmark for Evaluating Nuclear Domain Competence in Large Language Models
arXiv cs.AI, 2026-11-18.
A Temporal Algebraic Framework for Data Stream Evaluation Under Delayed Labels and Distribution Shifts
arXiv cs.AI, 2026-11-18.

Security, offense, and the attack surface

Stopping Is Not the Limit: Understanding and Eliciting Agent Capability in Long-Horizon Vulnerability Exploitation
arXiv cs.AI, 2026-11-18. The capability ceiling in exploitation is elicitation, not ability. Follows directly from yesterday's FBI and Google reports of agents already doing this in the field.
Adaptive LLM-Driven Honeypots vs. LLM-Driven Attackers: Does Narrative Coherence Improve Deception?
arXiv cs.AI, 2026-11-18.
Optimizely Hack Turns World's Biggest Websites Into Malware Dispensers
Slashdot, 2026-11-17. A supply-chain compromise in an A/B testing script, which runs before everything else on the page by design.
Critical Red Hat Quay Flaw Lets Attackers Read and Overwrite All Private Container Images in a Registry
Lobsters, 2026-11-17.
Java4Shell: Write once, RCE everywhere
Pinboard, 2026-11-18.
Phrack 73
Lobsters, 2026-11-17.
Uber Says Hackers Stole Info On Nearly 150K Driver's Licenses During Data Breach
Slashdot, 2026-11-17.
'123456' Has Been the World's Most Popular Password for 7 of the Last 8 Years
Slashdot, 2026-11-17.
Formal Verification of the Signal Protocol Using AI
Schneier on Security, 2026-11-17. Machine-assisted proof applied to a protocol people actually depend on. The constructive end of the same capability that produces the exploitation results above.
Network-Exposed Configuration Attacks on LLM-Controlled Social Robots
arXiv cs.AI, 2026-11-18.
LLMs as Validators for Cross-Chain Bridges: Evaluating Source-Level and Transaction-Level Attack Detection on Solana
arXiv cs.AI, 2026-11-18.
From Ideal to Real: Integrating Language Models into Smart Home for Human-Centric IoT Safety and Security Fuzzing
arXiv cs.AI, 2026-11-18.
Custom Firmware for Fortinet FortiGate 60 Firewall/Router
Hackaday, 2026-11-17.

Infrastructure, outages, and the buildout

Apple services, including iCloud, App Store, Apple Music, and more, hit by outage
TechCrunch, 2026-11-17.
Highly Classified NSA Network Goes Down
Slashdot, 2026-11-17.
Google joins the tech outage trend by announcing it's not having one
TechCrunch, 2026-11-17. Issuing a statement that you are up is a new genre, and the fact that it was warranted says more than the statement.
Codeberg's domain suddenly disappeared from their account, site is unreachable
Lobsters, 2026-11-18.
The day the sky emptied: rendering the FAA outage with ClickHouse FORMAT PNG
ClickHouse, 2026-11-18. An outage rendered straight out of the database as an image. Both a good analysis and a reminder of how much the FAA disruption is still being worked through.
Quad9's withdrawal from France: A canary in the coal mine
Cloudflare, 2026-11-17.
Microsoft's ARM-based Cobalt 200 chips go GA. Why it matters for AI agents.
The New Stack, 2026-11-17.
CoreWeave connects custom training loops to model serving
The New Stack, 2026-11-17.
Solver-Free and Cost-Aware Online Scheduling of AI Data Center Tasks
arXiv cs.AI, 2026-11-18.
HyCoInfer: Cloud-Edge Collaborative Inference for Large Language Models on Heterogeneous GPU-FPGA Architectures
arXiv cs.AI, 2026-11-18.
SLED: Sensor SLicing from Edge to Datacenter
arXiv cs.AI, 2026-11-18.
As U.S. voters reject data centers, tech companies look to Spain
WBUR, 2026-11-17. The buildout relocating rather than slowing. Local referenda move the siting problem across a border.
Kraft-backed data center proposal met with opposition from Sutton residents
WBUR, 2026-11-17.
Hydropower hit record lows in the US Southwest this summer
MIT Technology Review, 2026-11-17.
Deep Reinforcement Learning for Heat Pump and Underfloor Heating Control: An End-to-End Case Study
arXiv cs.AI, 2026-11-18.
When does deep reinforcement learning pay off in HVAC supervisory control? Four small Modelica case studies trained directly on FMUs
arXiv cs.AI, 2026-11-18.
China dominates battery graphite, but this startup could break the monopoly
TechCrunch, 2026-11-18.

Data platforms and analytics

DuckDB 2.0 goes client-server and adds an Agent Mode for coding agents
The New Stack, 2026-11-16. An embedded database growing a server and an agent interface in the same release. Both additions answer the same pressure.
DuckDB User Group Workshop on AI-Powered Review Analysis with Hugging Face
DuckDB, 2026-11-17.
ClickHouse forecasting and anomaly detection with time series foundation models
ClickHouse, 2026-11-17.
More load is faster: Tuning Postgres read replicas
ClickHouse, 2026-11-17. A counterintuitive result worked through carefully. Replica behavior under load is not monotonic.
How we raised Databricks cluster rate limits by 7.5x
Databricks, 2026-11-18.
Built on object storage, served from cache
Databricks, 2026-11-17.
Advancing Catalog Interoperability with Unity Catalog and Google Cloud's Borderless Lakehouse
Databricks, 2026-11-17.
Branch, sync, serve: What our Lakebase Hackathon taught us about building on Lakebase
Databricks, 2026-11-17.
How packaged clean rooms bring RampID to your lakehouse
Databricks, 2026-11-17.
Find the log line that matters faster with full-text search in Grafana Cloud Logs
Grafana Labs, 2026-11-17.
Redict deprecated in favor of Valkey
LWN, 2026-11-17. One Redis fork folding into the other. The post-relicensing consolidation reaching its predictable end.
EviDARL: Evidence-Centric Dual-Agent Reasoning via Alternating Reinforcement Learning for Multi-Table Question Answering
arXiv cs.AI, 2026-11-18.
Efficient Wide-Beam Generative Retrieval with Model-Native Trie Decoding and Shared-Context Attention
arXiv cs.AI, 2026-11-18.

Languages, compilers, and the craft

Four major changes coming to Rust
InfoWorld, 2026-11-18.
Actual curl rewrites in rust
Lobsters, 2026-11-17. Stenberg distinguishing the incremental Rust adoption actually happening in curl from the rewrite people keep announcing on his behalf.
How to Bypass Rust's Orphan Rule
Lobsters, 2026-11-17.
C's designated initializers give us named arguments
Lobsters, 2026-11-17.
Automatic Checking of Compiler Correctness with E-Graphs
Lobsters, 2026-11-17.
LZ-RDO lands in ARM's astc-encoder
Lobsters, 2026-11-18.
Zellij manages terminal workspaces with style
LWN, 2026-11-17.
Running the full server-side Datasette Python app in the browser with service workers and Pyodide
Simon Willison, 2026-11-17. A server-side Python app running entirely client-side via service workers and Pyodide. The deployment story for small tools keeps collapsing toward "it is a web page."
FreeBSD GUS
Lobsters, 2026-11-17.
Choosing the Right Disks for OpenZFS: HDD, SSD, NVMe
Klara Systems, 2026-11-17.
Three stable kernels for Tuesday
LWN, 2026-11-17.
Security updates for Tuesday
LWN, 2026-11-17.
OCaml Weekly News, 17 Nov 2026
OCaml.org, 2026-11-17.
Sarasa Gothic Mono
Lobsters, 2026-11-17.

Clojure and the Lisps

SCI sandboxing on JS host considered unsafe
Planet Clojure, 2026-11-17.
pumice 0.2.0
Planet Clojure, 2026-11-16.
A Shared UI Convention for Clojure (Part 3)
Planet Clojure, 2026-11-16.
Sexp Matrix
Lobsters, 2026-11-17.
My Emacs 31 contributions
Lobsters, 2026-11-17.

Robotics, world models, and embodiment

When the World Is the Test Suite: Agentic Intelligence for the Physical World
arXiv cs.AI, 2026-11-18.
RoboTeam: Benchmarking Vision-Based Multi-Agent Control of Many Robot Arms in Real Time
arXiv cs.AI, 2026-11-18.
FailureGen: Scalable Robot Failure Data Generation with World Models
arXiv cs.AI, 2026-11-18.
Testing World Models on Actions They Have Never Seen: Held-Out Plans, Forecast Accuracy and Price Rules in Retail
arXiv cs.AI, 2026-11-18. Held-out actions rather than held-out time. The right generalization test for a world model, and rarely the one reported.
Towards Data-Efficient, Real-Time World Models
arXiv cs.AI, 2026-11-18.
Perception into Imagination: Executable Memory for World Models
arXiv cs.AI, 2026-11-18.
Manufacturing Differential Sparsity: Planner-Hardware Co-Design for Efficient World Model Planning
arXiv cs.AI, 2026-11-18.
Learning Agent-Centric Vision-Language-Action World Models for Autonomous Driving
arXiv cs.AI, 2026-11-18.
AutoE2E: Don't speak - Language free end-to-end autonomy. A Real-Time Driving Model for L2++ ADAS and L4 Robotaxis
arXiv cs.AI, 2026-11-18.
FIRE-Drive: Feasibility-Informed Risk Evaluation for End-to-End Diffusion Planning
arXiv cs.AI, 2026-11-18.
Reconfiguring Procedural Knowledge for Compositional Robot Skill Adaptation
arXiv cs.AI, 2026-11-18.
MillBill: A Machining Agent with Geometric Feedback
arXiv cs.AI, 2026-11-18.
Big Brains, Tiny Board: Pupavision, an Agentic Physical AI Instrument for Monitoring Drosophila Pupariation and Lifecycle Transitions
arXiv cs.AI, 2026-11-18.
Waymo suspends freeway rides nationwide after California DMV order
TechCrunch, 2026-11-17. A regulator ordering a capability off the road rather than negotiating over it. The same instrument the state used against OpenAI on Monday.
Split-and-Scale Latent 3D Representations
Apple ML Research, 2026-11-16.

Clinical and life sciences

Right Diagnosis, Unsafe Route: Auditing Clinical AI Decision Trajectories and Corrective Oversight for Validity and Potential Harm
arXiv cs.AI, 2026-11-18. The correct answer reached by an unsafe path. Outcome-only evaluation cannot see this, and clinical deployment is where it matters most.
Not All Evidence Is Admissible: Evidence Governance for Tool-Augmented Medical Agents
arXiv cs.AI, 2026-11-18.
DiagTrace: An Interactive Benchmark for Clinical Decision Making
arXiv cs.AI, 2026-11-18.
OICycle: Observe, Infer, and Intervene Cycle for Belief-State Learning in Longitudinal Clinical Decision Making
arXiv cs.AI, 2026-11-18.
MedGuardBench: A Patient-Oriented Benchmark for Personalized Guardrail Development
arXiv cs.AI, 2026-11-18.
VITAL: A Framework for Validating Agentic AI in Pharmaceutical Drug Development and Clinical Trial Operations
arXiv cs.AI, 2026-11-18.
OAK-Mind: Zero-Training Mental Health Change Inference from Passive Sensing via Knowledge-Grounded Multi-Agent LLMs
arXiv cs.AI, 2026-11-18.
EDA: Observe, Verify, and Assess for Evidence-Grounded Multimodal Depression Analysis
arXiv cs.AI, 2026-11-18.
Fine-Grained Japanese Clinical Named Entity Recognition: A Controlled Comparison of BERT and Open-Weight Large Language Models
arXiv cs.AI, 2026-11-18.
US Officially a Measles Endemic Country
Slashdot, 2026-11-18.
A bleeding disorder, once nearly vanquished, is on the rise
WBUR, 2026-11-18.
This smell startup says its AI nose is as good as a human's
MIT Technology Review, 2026-11-18.
Can AI help build a healthier future?
Microsoft Research, 2026-11-17.

Policy, society, and the human read

AI models can be tricked into identifying civilians as military targets by telling them it's a game
TechCrunch, 2026-11-17. Reframing the task as play dissolves the refusal. The most consequential jailbreak class and the least technical.
The Department of Justice Is Using AI Detector Pangram Against Defense Attorneys
404 Media, 2026-11-17. An unvalidated classifier used as an accusation in court. Detector error rates stop being an academic matter at this point.
'Defense by ChatGPT'
404 Media, 2026-11-17.
Google To Keep Temporary Gemini Chats For Two Months
Slashdot, 2026-11-17.
Google to keep Gemini chats for 2 months, up from 72 hours, even with history off
TechCrunch, 2026-11-17. Retention extended by a factor of twenty for conversations the user marked temporary. "History off" now means something different than it did.
Massive New Study Shows How Young People Feel About AI Companions
404 Media, 2026-11-18.
A Man and His Muse AI Agent Sue Meta
404 Media, 2026-11-17.
Voters Reject ALPRs and Surveillance Tech at the Ballot Box
EFF Deeplinks, 2026-11-18.
European Commission Plan to Give Frontex Trusted Flagger Status Is A Threat to Free Speech
EFF Deeplinks, 2026-11-17.
The Justice Department Wants To Give Patent Trolls More Power
EFF Deeplinks, 2026-11-17.
The Paramount-Warner Brothers Deal Could Usher in Another Age of Lost Media
EFF Deeplinks, 2026-11-17.
Enshittifiers groom us for scammers
Pluralistic, 2026-11-18.
Honest Brokers in AI Regulation: How Public Comments Introduce Evidence to Regulators
arXiv cs.AI, 2026-11-18.
Can Power Draw Constrain Covert Compute? Limits of Analogue Verification for AI Governance
arXiv cs.AI, 2026-11-18.
Measuring Human-AI Alignment as an Early Indicator for Effective Collaboration in Hybrid SE Teams
arXiv cs.AI, 2026-11-18.
Can AI Help Explain AI? Effects of Multimodal LLM-Generated Explanations on Perceived Understanding and Perceived Trust in XAI Visualizations
arXiv cs.AI, 2026-11-18.
School Phone Bans Are Gradually Improving Test Scores
Slashdot, 2026-11-17.
Could programs like Alcoholics Anonymous help with phone addiction?
WBUR, 2026-11-17.
The man inside ….
Kenneth Payne, 2026-11-17.
AI agents will soon surpass us
InfoWorld, 2026-11-18.

The crawler zoo

Twenty-four new user agents catalogued across two days, sixteen of them below. The names are the interesting part: AIAgentVulnScanner scanning for agent vulnerabilities, GeneralRetrievalBot and ResearchBot claiming the generic end of the namespace, xAI-SearchBot and DeepSeekBot arriving as named lab crawlers.

New arrival: AIAgentVulnScanner
The Crawler Zoo, 2026-11-18.
New arrival: GeneralRetrievalBot
The Crawler Zoo, 2026-11-18.
New arrival: ResearchBot
The Crawler Zoo, 2026-11-18.
New arrival: xAI-SearchBot
The Crawler Zoo, 2026-11-17.
New arrival: DeepSeekBot
The Crawler Zoo, 2026-11-17.
New arrival: BlueSkyAggregator
The Crawler Zoo, 2026-11-18.
New arrival: mwmbl-crawler
The Crawler Zoo, 2026-11-18.
New arrival: RankSharkBot
The Crawler Zoo, 2026-11-18.
New arrival: wheeler
The Crawler Zoo, 2026-11-18.
New arrival: Drupalbot
The Crawler Zoo, 2026-11-17.
New arrival: DataSeekerBot
The Crawler Zoo, 2026-11-17.
New arrival: OrgaMazeSpider
The Crawler Zoo, 2026-11-17.
New arrival: libwww-perl script
The Crawler Zoo, 2026-11-17.
New arrival: FavicoHub
The Crawler Zoo, 2026-11-17.
New arrival: NewsFeed
The Crawler Zoo, 2026-11-17.
New arrival: Catstodon
The Crawler Zoo, 2026-11-17.

Tail

Feed silences and verification

One hundred eighteen feeds polled this morning: 61 returned content, 56 answered 304, one errored.

arXiv cs.AI returned 401 items, 305 of them new and dated today, against 529 yesterday. The selector matched every item in the feed. C-004 remains unrefuted.

Hacker News returned HTTP 419 again, its twenty-third consecutive poll with that status, dating to 2026-10-27. The bug we disclosed on 2026-11-15 is still present: our conditional-fetch path classifies any response without changed content as :not-modified and logs the real status parenthetically, so the warehouse records a cache hit where the logs record a rejection. Hacker News last returned content on 2026-10-26 and contributes nothing to this brief. The fix is unfiled, because our issue tracker is not reachable from this host, so it is recorded here instead.

www.bornholt.dev is the one error. It has now failed to resolve on 176 consecutive polls since 2026-05-26 and has never once succeeded from this host. It is excluded from the publishing count rather than reported as a silence, because there is no evidence it was ever live from here.

One hundred thirty-two items in the window carry source = 'wal.sh research' in the warehouse. That is not a source, it is us: our own Atom feed is crawled back in, url is the primary key, and re-ingestion overwrites the original publisher on every item we have previously curated. Those items are attributed above by their URL domain rather than by the stored source field. The count has gone 113 (11-14) → 125 (11-15) → 126 (11-16) → 116 (11-17) → 132 (11-18). Yesterday's drop was the window rolling past older briefs; today's rise is yesterday's brief being ingested. The fix belongs in the store.

build: 2026-11-18 | crawler-sha: b277e41 (Walsh-Research/1.2, compliance v1.3) | feeds: 118 polled, 70 publishing | items-considered: 5,822 (14d, incl. 3,558 arXiv) | warehouse: 72,977 items | published: PENDING