Morning Brief: Friday, November 20

Seventy-four feeds reporting. Two weeks. 5,829 items reduced to what follows. (what we track, how we crawl, subscribe)

Agent memory and agent skills both arrived this morning as software artifacts – with drift detectors, defect taxonomies and repair costs – on the same day a paper argued that agent safety cannot be verified at all.

The memory cluster is the larger of the two, and its titles are operational rather than architectural. How Much Should an Agent Remember? frames retention as a deployment parameter with measured tradeoffs. DriftGuard detects memory drift online via causal attribution, which presumes drift is a defect class you monitor for in production. MINTEval evaluates memory under multi-target interference; CompAct-Bench benchmarks working-context compaction; Forget Freely, Grep Later argues for index-free memory on the theory that grep at read time beats a structure you have to maintain. The skill papers use the same vocabulary one layer up: SkillBridge diagnoses and repairs cross-OS compatibility defects in agent skills and prices the repair, Behavioral Integrity Verification for AI Agent Skills treats a skill as something whose behavior you attest, and EvoSkill Injection red-teams the generation pipeline itself.

Against that, Pandora's Toolbox Or Why Agent Safety Cannot Be Verified is the morning's negative result, and it did not arrive alone: Tether ships a verified stateful policy engine for runtime monitoring regardless, and The Price of an Explanation tries to quantify what demanding verifiable artifacts costs and buys. The empirical companion is Characterizing and Detecting Unfaithful Work Reports by Coding Agents – agents that misdescribe what they did, which is the failure mode that makes verification load-bearing rather than ceremonial. The New Stack reported yesterday that a Devin deployment went 22 of 44 days without a human message and shipped an Android app anyway.

Top (5-7 min)

Pandora's Toolbox Or Why Agent Safety Cannot Be Verified
arXiv cs.AI, 2026-11-20. A negative result on the thing the rest of the stack is being built to deliver. Worth reading next to the papers below that ship verification machinery anyway.
Characterizing and Detecting Unfaithful Work Reports by Coding Agents
arXiv cs.AI, 2026-11-20. Agents describing work they did not do, treated as a detectable phenomenon with characteristics. This is the concrete case for why the verification layer is not optional.
SkillBridge: Cost-Aware Diagnosis and Repair of Cross-OS Compatibility Defects in Agent Skills
arXiv cs.AI, 2026-11-20. Agent skills get a defect class, a diagnosis step, and a repair cost. The framing is maintenance engineering, not prompt engineering.
How Much Should an Agent Remember? Deployment Lessons from Self-Distilled Memory for LLM Agents
arXiv cs.AI, 2026-11-20. Retention as a tunable with measured consequences, reported from deployment rather than from a benchmark.
Anthropic Tops $2 Trillion Market Cap After IPO Debut
Slashdot, 2026-11-19. Priced at $83 on a $1.5T valuation, opened at $111, topped $2T market cap the same session. TechCrunch reports the seven founders took out nearly $2B and that the structure could route $156B to charity.
A single pull request could have handed attackers the keys to Microsoft's software supply chain
InfoWorld, 2026-11-20. Second Microsoft supply-chain story in three days, after the agent-framework package hijack that stole Azure and CI credentials.
AI coding agents need a safe place to work; Eclipse is building the sandbox to give it to them
InfoWorld, 2026-11-20. Containment as foundation-level infrastructure, arriving the same week OpenAI published the internals of the Codex sandbox.

Themes

Memory is a deployment parameter now

Ten papers this morning treat agent memory as something you size, monitor, benchmark and repair rather than something you architect once. The shift in vocabulary – drift, interference, compaction, eviction, revision – is the vocabulary of operating a system in production.

How Much Should an Agent Remember? Deployment Lessons from Self-Distilled Memory for LLM Agents
arXiv cs.AI, 2026-11-20. Retention framed as a cost/benefit dial, with deployment lessons rather than a leaderboard.
DriftGuard: Online Memory-Drift Detection via Causal Attribution for LLM Agents
arXiv cs.AI, 2026-11-20. Memory drift as a production defect you detect online, attributed causally rather than flagged by heuristic.
Forget Freely, Grep Later: Lossless, Index-Free Memory for Long-Horizon LLM Agents
arXiv cs.AI, 2026-11-20. The anti-structure position: keep everything flat and search at read time rather than maintain an index that can go stale.
MINTEval: Evaluating Memory under Multi-Target Interference in Long-Horizon Agent Systems
arXiv cs.AI, 2026-11-20. Interference between remembered targets as the thing being measured, which is a harder test than single-thread recall.
CompAct-Bench: A Benchmark for Working-Context Compaction in Long-Horizon Agent Tasks
arXiv cs.AI, 2026-11-20. Compaction gets a benchmark. Anyone who has watched a long agent session summarize itself badly knows why this needs one.
HabitusMem: Belief Memory over Evolving User Preferences for LLM Agents
arXiv cs.AI, 2026-11-20. Memory of beliefs about a user, with the user's preferences treated as a moving target.
UpliftMem: Learning Set-Level Uplift for Agent Memory Retrieval
arXiv cs.AI, 2026-11-20. Retrieval scored by what the retrieved set actually improves, not by similarity.
STARK: Moving towards Personalized Agents That Truly Learn and Remember
arXiv cs.AI, 2026-11-20.
System-One RSI: Building Muscle Memory for Interactive Agents through Guided Practice
arXiv cs.AI, 2026-11-20. Practice-acquired reflex rather than retrieved recall – a different sense of memory than the rest of the cluster.
MemMend: A Selective Revision Framework for Memory Question Answering
arXiv cs.AI, 2026-11-20.
Global Importance Estimation for KV Cache Eviction
arXiv cs.AI, 2026-11-20. The same problem one layer down, in the cache rather than the agent.

Skills are software, and software has defects

Five papers treat an agent skill the way you would treat a package: something with a compatibility matrix, an integrity attestation, a coverage story, and an injection surface.

SkillBridge: Cost-Aware Diagnosis and Repair of Cross-OS Compatibility Defects in Agent Skills
arXiv cs.AI, 2026-11-20.
Behavioral Integrity Verification for AI Agent Skills
arXiv cs.AI, 2026-11-20. Attesting what a skill does, which only becomes necessary once skills are distributed artifacts rather than local prompts.
EvoSkill Injection: Red-Teaming Autonomous Skill Generation and Evolution in Self-Evolving Agents
arXiv cs.AI, 2026-11-20. If an agent writes its own skills, the generation pipeline is an attack surface. This red-teams it.
Evolve Where It Matters: Coverage-Guided Self-Evolution of Agent Skills
arXiv cs.AI, 2026-11-20. Coverage-guided evolution, borrowing the idea straight from fuzzing.
How Feedback-Free Self-Evolution Reshapes Agent Skills: A Systematic Study
arXiv cs.AI, 2026-11-20. What happens to a skill set that evolves with no external signal telling it whether it got better.
Demonstration-Free Skill Composition for Navigation and Manipulation
arXiv cs.AI, 2026-11-20. The embodied version of the same question.

What verification can and cannot buy

Pandora's Toolbox Or Why Agent Safety Cannot Be Verified
arXiv cs.AI, 2026-11-20.
Tether: Verified Stateful Policy Engine for System-Level Monitoring of Agents
arXiv cs.AI, 2026-11-20. Verify the monitor instead of the agent. A reasonable response to the impossibility result above.
The Price of an Explanation: What Demanding Verifiable Artifacts Costs, and What It Buys Back
arXiv cs.AI, 2026-11-20. Verification treated as a purchase with a price, which is the honest framing once you accept it is not free.
Verify the Verifier: Specification-Grounded LLM Cross-Checking for Bugs That Escape UVM Scoreboards
arXiv cs.AI, 2026-11-20.
Anchor and Check: Deterministic Cross-Prompt, Cross-Model Reconciliation for Auditable LLM Decisions
arXiv cs.AI, 2026-11-20.
Backtracking Agent Applications Security Incidents
arXiv cs.AI, 2026-11-20. Post-incident forensics for agent applications, which implies incidents are now frequent enough to need a method.
Replication of Agentic Security Events: Audit and Investigation (ROGUE-AI)
arXiv cs.AI, 2026-11-20. Replaying an agentic security event so it can be investigated. The pair with the above is a nascent incident-response discipline.
Why verification, not generation, is your secret to AI coding ROI
The New Stack, 2026-11-18.

Scan

Agent memory, skills, and harnesses

Beyond Weight Updates: Evolving Agent Harnesses for Domain Specialization
arXiv cs.AI, 2026-11-20.
Harness-Zero: Harness Distillation via Agent-as-Harness
arXiv cs.AI, 2026-11-20.
Meta-RTL: Evolving Agent Harnesses for RTL Optimization
arXiv cs.AI, 2026-11-20.
CANDID: Candidate-Driven Intervention for Disciplined Execution in Local Computer-Use Agents
arXiv cs.AI, 2026-11-20.
Stop Going in Circles: Loop-Triggered Routing for System 1 Decision Agents
arXiv cs.AI, 2026-11-20.
MATE: Mobile Agent Trustworthy Execution for Safe and Efficient Mobile Control
arXiv cs.AI, 2026-11-20.
Tokengeist: Multi-Turn Attribution Tracing in Agentic Conversations
arXiv cs.AI, 2026-11-20.
Sage-Evolve-Preview: Adaptive Observation for Post-Training Self-Improvement
arXiv cs.AI, 2026-11-20.
Validation-driven Deployment of LLM Inference Platforms for Agentic Workloads
arXiv cs.AI, 2026-11-20.
FinVeriQA: Financial Research Agents for Question Answering: Routing, Retrieval, MCP Tool Use, and Provenance-Aware Evidence Verification
arXiv cs.AI, 2026-11-20.
Generalist Meets Specialists: An Evolving Multimodal Agent with Persistent Tool Orchestration for Traffic Scenario Understanding
arXiv cs.AI, 2026-11-20.
Seer: Guiding Small Deep-Research Agents with Shared Decision Models
arXiv cs.AI, 2026-11-20.
Does your AI agent need an LLM to decide what to do next?
The New Stack, 2026-11-20. AWS argues a smaller decider model can do the routing, which is the same cost-surface argument the harness papers are making.
A "headless Slack": Vercel's new message board lets Claude Code and Codex agents work together
The New Stack, 2026-11-19.
An engineer didn't message his AI for 22 of 44 days. It built an Android app anyway.
The New Stack, 2026-11-19.
"Be ready for change": AAIF's CTO on why the agent stack won't settle
The New Stack, 2026-11-19.
IDrive wants to back up your coding agents, too
The New Stack, 2026-11-19.
Agentic AI repeats same old enterprise adoption cycle
InfoWorld, 2026-11-20.
AI agents will soon surpass us
InfoWorld, 2026-11-18.

Software engineering and coding agents

Characterizing and Detecting Unfaithful Work Reports by Coding Agents
arXiv cs.AI, 2026-11-20.
Copilot or Culprit? How AI Is Changing Open-Source Bug Reporting and Maintenance
arXiv cs.AI, 2026-11-20. Maintainer-side measurement of what agent-authored bug reports are doing to open source, which is the part of the story the vendors do not publish.
BuggyMLE: Scaling Machine Learning Engineering Difficulty with Bug Injection
arXiv cs.AI, 2026-11-20.
AppSci-Bench: Can LLMs Build AI Agents and Make Them Better?
arXiv cs.AI, 2026-11-20.
AutoRAGBench: A Benchmark for LLM Agents Automating RAG System Development
arXiv cs.AI, 2026-11-20.
Agent^2 RL-Bench: Can LLM Agents Engineer Agentic RL Post-Training?
arXiv cs.AI, 2026-11-20.
DVBench: End-to-End Benchmark for Real-World UVM Design Verification Agents
arXiv cs.AI, 2026-11-20.
Trace2ST: Recovering PLC Structured Text from I/O Traces with LLM Generation and Verification
arXiv cs.AI, 2026-11-20.
DatalogBench: Evaluating Large Language Models on Text-to-Datalog Synthesis
arXiv cs.AI, 2026-11-20.
EverydayBench: Evaluating Satisfactory Completion of Live User-Facing Digital Work
arXiv cs.AI, 2026-11-20.
DSFlow: A Neuro-Symbolic Data Workflow Agent with Statistical Diagnostic Guardrails
arXiv cs.AI, 2026-11-20.
Reverse-Engineering Kevin J. Lang's Lost OCaml Code in Apache DataSketches
Pinboard, 2026-11-19. A concrete, bounded repair task done with a model, written up by the person who did it. More useful than most benchmark papers on the same question.
Tracking Down a Mystery Beep By Giving a Coding Agent a Few Ears
Hackaday, 2026-11-19.
SDR# VDK: Create Your Own SDR# Plugins by Prompting an AI Coding Agent
RTL-SDR, 2026-11-20.
Token fraud is here and it's BAD
Simon Willison, 2026-11-20. Pairs with the unfaithful-work-reports paper above: billing surface and reporting surface are the same surface.
Claude Code v2.1.329
claude-code releases, 2026-11-19.
"I can tell you wrote this yourself"
Lobsters, 2026-11-20.
My Least Favorite Example in Software Engineering
Hillel Wayne, 2026-11-19.
Design for Understandability
Lobsters, 2026-11-19.
Data Contracts
Martin Fowler, 2026-11-18.
Stacked diffs with Maiao
Lobsters, 2026-11-18.
Announcing jj-musubi: live sync for Jujutsu
Lobsters, 2026-11-20.
Jujutsu Primer
Lobsters, 2026-11-19.
Accelerating development for massive Git repositories
Lobsters, 2026-11-18.
The best things in open source are built together
GitHub Blog, 2026-11-19.

Evaluation, benchmarks, and interpretability

Blind Spots Travel, Knowledge Stays Home: What RL Inherits from an LLM Judge
arXiv cs.AI, 2026-11-20. What a judge model's failures do to the policy trained against it. A judge is a dependency, and this treats it as one.
AltRA-Test: An Alternative Annotator Test for LLM Judges Without Shared Annotators
arXiv cs.AI, 2026-11-20.
Calibrated Decision Rewards: Training Language Models with a Judge That Knows How Sure It Is
arXiv cs.AI, 2026-11-20.
LLM GerryMandering: Separating Interactions from Additive Effects in LLM Audits
arXiv cs.AI, 2026-11-20.
The Geometry of Lies: Knowledge and Persona Directions for Detecting Deception in Large Language Models
arXiv cs.AI, 2026-11-20.
Does the Workspace Do the Work? A Causal Protocol for Verifying Global-Workspace Computation in a Language Model
arXiv cs.AI, 2026-11-20.
Mechanistic Interpretability Needs Philosophy
arXiv cs.AI, 2026-11-20.
MambaLIFT: Reference-Based Mechanism-Aware Attribution for VMamba
arXiv cs.AI, 2026-11-20.
Many Solutions, One Answer: Search Order Shapes the Default Solutions of Language Models
arXiv cs.AI, 2026-11-20.
ClosureBench: A Benchmark for Reasoning under Closure Assumptions
arXiv cs.AI, 2026-11-20.
PosteriorBench: Evaluating Behavior Posterior Sampling in Language Models
arXiv cs.AI, 2026-11-20.
One Framework to Diagnose Them All: A Stratified Benchmark for Bengali LLM Hallucination
arXiv cs.AI, 2026-11-20.
Diagnosis is not a policy: a state-transition audit of capability-conditioned test-time inference
arXiv cs.AI, 2026-11-20.
A Model of Diverse Sampling from Language Models
arXiv cs.AI, 2026-11-20.
LLM Complexity: Towards a Computational Complexity View of Large Language Model Inference
arXiv cs.AI, 2026-11-20.
Rigel-Ladon: An Option-Count Law for a One-Pass Decision Encoder, Tested on Held-Out Option Counts and Held-Out Schemas
arXiv cs.AI, 2026-11-20.
Sirius-Tiamat: How Small Can a One-Pass Decision Encoder Get Before Its Decisions Change?
arXiv cs.AI, 2026-11-20.
AI GameStore: Scalable, Open-Ended Evaluation of Machine General Intelligence with Human Games
arXiv cs.AI, 2026-11-20.

Security, offense, and the attack surface

nginx2root: Five new nginx vulnerabilities, and how Cloudflare protects origins from the worst of them
Cloudflare, 2026-11-20.
NGINX2Root: Remote Code Execution vulnerabilities in nginx
LWN, 2026-11-18.
nginx2root
Lobsters, 2026-11-18.
Java4Shell: Write once, RCE everywhere
Pinboard, 2026-11-18.
Java4Shell is a reason to use Java
Lobsters, 2026-11-18. The contrarian read: the thing that made the bug findable and patchable fast is the same thing people complain about.
How Cloudflare WAF helps protect against Java4Shell (CVE-2026-114938)
Cloudflare, 2026-11-18.
Java4Shell vulnerability (CVE-2026-114938) assessment
Databricks, 2026-11-19.
Ruby2Shell (GHSA-34m4-m9fh-gqp5)
Lobsters, 2026-11-19.
Security advisories: multiple vulnerabilities in Regexp and String case conversion
Lobsters, 2026-11-18.
Four PHP security releases including fixes for critical vulnerabilities
LWN, 2026-11-19.
Attackers hijack Microsoft's agent-framework package to steal Azure and CI credentials
InfoWorld, 2026-11-18.
The U.S. national cyber director is stepping down
TechCrunch, 2026-11-20. Departure lands in the middle of the outage-and-advisory stretch that has run through this whole window.
JPMorgan, Visa, Mastercard Disclose Cyberattack On Third-Party Vendors
Slashdot, 2026-11-19.
Bank CEOs, NSA Tell Congress That AI-Driven Patches Are Behind Outages
Slashdot, 2026-11-18.
Amidst fears of AI cyberattacks, Satya Nadella says lots of critical infrastructure is 'unpatchable'
TechCrunch, 2026-11-18.
Palmer Luckey warns AI-powered cyberattacks are outpacing defenses
TechCrunch, 2026-11-18.
SoK: Assessing the Evolution of LLM Capability in Software Vulnerability Research
arXiv cs.AI, 2026-11-20.
ChainAgent: Automated Chainable Vulnerability Discovery in PHP Web Applications
arXiv cs.AI, 2026-11-20.
From FIPS to FPGA with LLM Agents: Measuring What Tests and Reviews Miss in Post-Quantum Cryptographic Hardware
arXiv cs.AI, 2026-11-20.
A Wolf in Sheep's Pixels: Visual Camouflage Attacks via HDR Gain Maps
arXiv cs.AI, 2026-11-20.
ChronoPatch: Black-Box Energy-Latency Attacks on Vision-Language Models
arXiv cs.AI, 2026-11-20.
SteerTrace: Tracing How Target Attacks Steer World Action Models
arXiv cs.AI, 2026-11-20.
When Safety Rationales Hallucinate: Diagnosing and Grounding Multimodal Safety Guards
arXiv cs.AI, 2026-11-20.
CP-Guard+: A New Paradigm for Malicious Agent Detection and Defense in Collaborative Perception
arXiv cs.AI, 2026-11-20.
Crafting Targeted MP4 Fuzzers
Lobsters, 2026-11-20.
Agent security is product security, and your marketing site counts
The New Stack, 2026-11-19.
"The best escape room players that you've ever met": What Red Hat's AI architect really thinks about agents
The New Stack, 2026-11-18.
Inside the Codex sandbox
OpenAI, 2026-11-18.
Stricter checks for uploaded distributions on PyPI
Lobsters, 2026-11-18.
France Is Blaming Iran for the Fridge Cyberattack
Schneier on Security, 2026-11-18.
To Catch a Ransomware Scammer, FBI Posed as Both Victim and Attacker
404 Media, 2026-11-19.
UK government forced to disclose secret Apple backdoor order
TechCrunch, 2026-11-19.
FreeBSD: How Do You Tackle CVE Security Patch Cadence?
Klara Systems, 2026-11-19.
Supporting FOSS projects to transition to memory safe tech
Lobsters, 2026-11-19.

Infrastructure, inference, and the buildout

Budget Internalization: Fixed-Budget RLVR Trades Test-Time Scaling for In-Budget Accuracy
arXiv cs.AI, 2026-11-20.
The ROI of Small Examples: Optimizing In-Context Learning Under Token Budgets
arXiv cs.AI, 2026-11-20.
Cost-Aware Agentic Feature Optimization System for Fleet-Scale Ranking Models
arXiv cs.AI, 2026-11-20.
Splice-Core: Low-Order-Bit-Splicing GEMM for Highly Energy-Efficient AI Inference Accelerators
arXiv cs.AI, 2026-11-20.
Online Temporal Rotations for 2 Bit KV-Cache Quantization in Video Diffusion Transformers
arXiv cs.AI, 2026-11-20.
STEPQuant: When and Where Errors Matter in Delta-Rule Recurrent State Quantization
arXiv cs.AI, 2026-11-20.
Ask in the Crowd: Differentially Private LLM Inference via Dummy-Augmented Shuffling
arXiv cs.AI, 2026-11-20.
Asuka-FM: A System-Architectural Co-Design Framework for Efficient Text-to-Image Pretraining on Consumer Hardware
arXiv cs.AI, 2026-11-20.
How OpenAI's Performance Factory keeps ChatGPT fast
OpenAI, 2026-11-19.
Where should AI workloads run? Probably next to the data.
The New Stack, 2026-11-19.
AI needs fresh data. Vector indexing is getting in the way
The New Stack, 2026-11-18.
200 OK is dead as an AI agent health signal. Here's what comes next.
The New Stack, 2026-11-18.
How Amazon ECS handles blue/green, canary, and linear deployments
The New Stack, 2026-11-19.
Google Signs Deal With Blue Origin to Send Satellite Data Centers Into Orbit
Slashdot, 2026-11-20.
Black Neighborhoods Are Twice as Likely To House a Data Center, Study Finds
404 Media, 2026-11-19.
The Hardware Worked: FAA Account Points to Cutover Gone Wrong, Not Failure of Old Equipment
Leeham News, 2026-11-19. The most specific account of the FAA outage so far, and it points at the migration rather than the legacy system.
The day the sky emptied: rendering the FAA outage with ClickHouse FORMAT PNG
ClickHouse, 2026-11-18.
These tech outages are giving Y2K, AI agent style
TechCrunch, 2026-11-19.
Outage Chaos: Who Turned Off All The Computer?
EFF Deeplinks, 2026-11-19.
Incidents in the news
Lobsters, 2026-11-18.
Traffic mix matters
Marc Brooker, 2026-11-18.
15 million web servers on one VM - smoltcp, DPDK, AF_XDP
Lobsters, 2026-11-19.
Detect, act, keep the evidence: node reliability across three clouds
ClickHouse, 2026-11-19.

Clojure and the Lisps

Clojure 1.12.7
Planet Clojure, 2026-11-18.
clojure-1.12.7 release
Clojure releases, 2026-11-18.
Babashka v1.13.230
Babashka releases, 2026-11-19.
ClojureDesign Ep 121: Current Events
Planet Clojure, 2026-11-19.
Lisp Game Jam devlog: Polydraw
Spritely Institute, 2026-11-19.

Research, math, and formal methods

A CPU Inside an LLM: Zero-Gradient Compilation of Exact Circuits into Pretrained Weights
arXiv cs.AI, 2026-11-20. Compiling an exact circuit into weights with no gradient steps. If it holds up it is a different relationship between program and model than fine-tuning.
WeightCode: Installing New Capabilities into Neural Networks via Weight-Space Programming
arXiv cs.AI, 2026-11-20.
A Tensor-Algebraic No-Go Theorem for High-Dimensional Photonic Greenberger-Horne-Zeilinger States Discovered with AlphaProof Nexus
arXiv cs.AI, 2026-11-20.
Actor-Predictor Architecture for Reinforcement Learning with Formal Specification Objectives
arXiv cs.AI, 2026-11-20.
A new algorithmic approach for finding flat minima: critically damped Polyak stepsizes
arXiv cs.AI, 2026-11-20.
Flow Matching from Non-Gaussian Sources: A Theory of Efficient Sampling
arXiv cs.AI, 2026-11-20.
Pareto-Optimal and Strategyproof Data-Sharing Between Competitors
arXiv cs.AI, 2026-11-20.
The FrontierOR Challenge: Agentic Algorithm Design for Operations Research
arXiv cs.AI, 2026-11-20.
SailSearch: Agent-Assisted Discovery in Combinatorial Optimization
arXiv cs.AI, 2026-11-20.
Similarity-based OSF Logic for Flexible SPARQL Query Answering
arXiv cs.AI, 2026-11-20.
Validating a High-Performance Cloud Object Store with Lightweight Formal Methods
Lobsters, 2026-11-19.
TLAPS-Bench: Formally Proving Correctness Of Real-World System Specifications With TLA+ Proof System
Murat Demirbas, 2026-11-18.
Claude Fable 5.2 solves two open problems in the physics of black holes
Pinboard, 2026-11-19.
A game you cannot win unless you have a quantum computer
Pinboard, 2026-11-20.
How M.C. Escher's Art Creates Bizarre Mathematical Worlds
Quanta Magazine, 2026-11-18.

Robotics, world models, and embodiment

World Models Favor Learned Priors over Visual Evidence
arXiv cs.AI, 2026-11-20. A failure mode with teeth: the model trusts what it learned over what it is currently seeing.
Bradley-Terry Training Can Distort Outcome Probabilities of Initially Accurate Language World Models
arXiv cs.AI, 2026-11-20.
Fast-JEPA: Advancing the Success-Latency Frontier for Latent World Models
arXiv cs.AI, 2026-11-20.
LATER: Learning Action-Conditioned Temporal Evidence Retrieval for Predictive World Models
arXiv cs.AI, 2026-11-20.
Don't Lose the Future: Latent-Aware 4-Bit Quantization of World Action Models
arXiv cs.AI, 2026-11-20.
MAWAM: A Multi-Agent World-Action Model for Coordinated Manipulation
arXiv cs.AI, 2026-11-20.
Mind2Act World: A Diagnostic Benchmark for Embodied Mind, Physical Action, and Mind-Action Composition in Robotic Manipulation
arXiv cs.AI, 2026-11-20.
A$^2$-VLA: Adapting Vision-Language-Action Policies to New Tools through Agentic Affordance Discovery
arXiv cs.AI, 2026-11-20.
PGX: Physically Grounded Execution for Long-Horizon Humanoid Loco-Manipulation
arXiv cs.AI, 2026-11-20.
Awaken-HI: Scaling Humanoid Intelligence with Whole-Body Human Experience
arXiv cs.AI, 2026-11-20.
Controlling the Humanoid Robot H1-2 With a Simulated Biologically Plausible Network
arXiv cs.AI, 2026-11-20.
Collision-Free Is Not Enough: Passage-Conditioned Navigation for Wheeled-Legged Robots
arXiv cs.AI, 2026-11-20.
Bridging vision-language reasoning and learned control for legged navigation in the wild
arXiv cs.AI, 2026-11-20.
CoachSI: Self-Improvement of Robot Policies with a Language Coach
arXiv cs.AI, 2026-11-20.
TipDex-VTLA: Learning When to Refine Actions with Touch for Dexterous Manipulation
arXiv cs.AI, 2026-11-20.
FlashX: Scaling Robotic Experience Reuse for Embodied Agents with CBA-Based NAND Flash
arXiv cs.AI, 2026-11-20.
Plan your Self-Adaptation! Ontology-Mediated Task and Architecture Co-adaptation Planning for Robots
arXiv cs.AI, 2026-11-20.
FastRoverVLA: Fast Vision-Language-Action Execution for Off-Road Autonomous Driving
arXiv cs.AI, 2026-11-20.
AI agents can redesign engineering parts with more than a thousand CAD parameters and nonlinear simulations
arXiv cs.AI, 2026-11-20.
Waymo Robotaxis Pulled From Freeways Indefinitely
Slashdot, 2026-11-18.

Policy, society, and the human read

Toward a Science of AI Agent Societies
arXiv cs.AI, 2026-11-20.
The Coevolution and Coexistence of Humans and AI
arXiv cs.AI, 2026-11-20.
When Agents Play Differently: Localizing Human-Agent Behavioral Divergence with Conditional Distributional Testing
arXiv cs.AI, 2026-11-20.
Why Do LLMs Help with Scientific Misconduct Requests and What to Do About It?
arXiv cs.AI, 2026-11-20.
Twice-Told Critiques: Claim-Level Homogenization of Peer Review in the LLM Era
arXiv cs.AI, 2026-11-20. Measures review convergence rather than asserting it, which makes it the useful companion to the Nature piece below.
When AI-assisted criticism scales, peer review needs an obligation threshold
Nature Machine Intelligence, 2026-11-18.
Steering Automated Scientific Discovery via Audited Epistemic Exploration
arXiv cs.AI, 2026-11-20.
LINEAGE: A Benchmark for Empirical Lineage Retrieval in AI Scientific Literature
arXiv cs.AI, 2026-11-20.
CoMPASS: A Multi-Turn Benchmark for Measuring LLM Reinforcement of Parasocial Behavior
arXiv cs.AI, 2026-11-20.
Massive New Study Shows How Young People Feel About AI Companions
404 Media, 2026-11-18.
Murder, it wrote: Spilling the tea on a cozy mystery community
WBUR, 2026-11-20.
A mom and teen on navigating phone and social media limits
WBUR, 2026-11-20.
College Board Delivers Another AI Ultimatum: No AI, No AP CS Principles Course
Slashdot, 2026-11-19.
A Quasi-experimental Study of Conversational AI-supported Engagement in Large CS Courses
arXiv cs.AI, 2026-11-20.
ICE Is Paying AT&T to Dig Through a Mountain of Call Records
404 Media, 2026-11-19.
Voters Reject ALPRs and Surveillance Tech at the Ballot Box
EFF Deeplinks, 2026-11-18.
Enshittifiers groom us for scammers
Pluralistic, 2026-11-18.
Managing frontier AI risk: From training to deployment
OpenAI, 2026-11-18.
Sam Altman warns against a future where AI, not humans, calls the shots
TechCrunch, 2026-11-18.
Cambridge Dictionary Names 'Human' Its 2026 Word of the Year
Slashdot, 2026-11-18.
Amazon Fights the Delivery Protection Act With Homepage Banner, In-App Pop-Up
Slashdot, 2026-11-20.

Funding, products, and the business read

Anthropic hits Nasdaq exchanges at $111, 33% above IPO price
TechCrunch, 2026-11-19.
Anthropic prices initial public offering
Anthropic, 2026-11-18.
Anthropic's 7 founders cashed out nearly $2B in record-breaking IPO
TechCrunch, 2026-11-19.
Anthropic's IPO could deliver $156B to charity
TechCrunch, 2026-11-19.
Anthropic's IPO puts AI's circular financing boom to the test
TechCrunch, 2026-11-19.
At Nasdaq opening bell, Anthropic CEO recommits to pacing AI development
TechCrunch, 2026-11-19.
Anthropic begins trading on the Nasdaq
WBUR, 2026-11-19.
AINews: Anthropic completes largest IPO in history, valued at $1.5T
Latent Space, 2026-11-20.
Anthropic rival says companies can (mostly) skip the pricey models
TechCrunch, 2026-11-19.
Qwen4-Max vs. Claude Opus 5.5: The cheaper model costs you somewhere else
The New Stack, 2026-11-18.
Tokenomics Foundation: Making sense of what AI actually costs you
InfoWorld, 2026-11-18.
The unfolding Jensen Huang scandal: a timeline
TechCrunch, 2026-11-19.
Nvidia Approaches Arm Boss Rene Haas In Search For New CEO
Slashdot, 2026-11-19.
Nvidia Rejects Jensen Huang's Offer to Return as CEO
Slashdot, 2026-11-18.
Nvidia's startup stakes ballooned to $65B, but Huang's exit puts future deals in doubt
TechCrunch, 2026-11-18.
Jensen Huang's Fall Hits 'Open' Models In Boost to Anthropic and OpenAI
Slashdot, 2026-11-18.
Palo Alto Networks Signs $4.5 Billion Deal For Applied Compute
Slashdot, 2026-11-19.
Microsoft CEO Satya Nadella explains why he asked 60K employees to quit
TechCrunch, 2026-11-19.
Satya Nadella predicts AI agents will reshuffle the economics of the web
TechCrunch, 2026-11-18.
Satya Nadella Says He's AGI-Pilled Now
Slashdot, 2026-11-18.
More banks say customers using AI agents are on their own, even if they get hacked
TechCrunch, 2026-11-18. Liability allocation arriving before the verification story does.
Meta's AI agent Muse is mixing up conversations and could be exposing private data
TechCrunch, 2026-11-19.
Where Will Boeing Cut Commercial Workforce? White-Collar Union Still Doesn't Know
Leeham News, 2026-11-20.
How to forfeit one of copyright's best labor policies
Pluralistic, 2026-11-19.
Le Groupe Figaro and OpenAI partner to bring Le Figaro journalism to ChatGPT
OpenAI, 2026-11-18.

Hardware, radio, and the physical

FoxSDR Update: New Main Window, 89 Plugins and a Call for RTL-SDR Beta Testers
RTL-SDR, 2026-11-20.
HackerGadgets Releases a 4-in-1 RTL-SDR, LoRa, GPS and RTC HAT for the Raspberry Pi
RTL-SDR, 2026-11-20.
HaxaWave: Turning the Modem in an Old Qualcomm MSM8916 Phone into a 14.8 MS/s SoapySDR Receiver
RTL-SDR, 2026-11-19.
Flipper One Launches on Kickstarter: A Linux Multi-Tool with an Optional SDR Module and MT7921 Wi-Fi
RTL-SDR, 2026-11-19.
Live Mesh Map: A Self-Hosted Live Map for Meshtastic and MeshCore LoRa Mesh Networks
RTL-SDR, 2026-11-19.
WaveSentry: Detecting Drone Signals with a HackRF, a Neural Network and an ESP32 Handheld
RTL-SDR, 2026-11-18.
Crashing Microcontrollers Competitively
Hackaday, 2026-11-20.
Preventing a Garage Door Opener From Rebooting When Offline
Hackaday, 2026-11-20.
Garage Door Spam
Lobsters, 2026-11-19.
DabbleCAD Takes Design Offline
Hackaday, 2026-11-20.
Serial: The Port That Won't Die
Hackaday, 2026-11-18.
Not-So-Large Language Model Fits On a Floppy Disk
Hackaday, 2026-11-18.
Monitors that Throttle Downloads
Hackaday, 2026-11-19.
One Hundred Gratuitous Case Fans? We're Fans of This.
Hackaday, 2026-11-20.
Rebuilding my voice assistant pipeline
Pinboard, 2026-11-20.
Advancing chip design with OpenAI models
OpenAI, 2026-11-18.
SiliconGym: A Versatile Reinforcement Learning Environment for Agentic EDA
arXiv cs.AI, 2026-11-20.
ReLay: Bidirectional Layout-Code Translation and Library Evolution for Reusable and Process-Portable Analog Design
arXiv cs.AI, 2026-11-20.
Satellites Now Outnumber the Visible Stars Above Most Towns
Slashdot, 2026-11-19.
NASA's Voyager 1 is so far away it now takes one day for its messages to reach Earth
WBUR, 2026-11-18.

The crawler zoo

Five new arrivals this morning, and two of them are the same idea spelled in both orders.

mcp-recon
The Crawler Zoo, 2026-11-20.
recon-mcp
The Crawler Zoo, 2026-11-20. Separate agents, separate operators, same reconnaissance target. The naming collision is itself the signal about how crowded this niche got.
catalog-health-check
The Crawler Zoo, 2026-11-20.
ant-html
The Crawler Zoo, 2026-11-20.
TLDWPTechBot
The Crawler Zoo, 2026-11-20.
mcp-auth-check
The Crawler Zoo, 2026-11-19.
AgentWebCensusBot
The Crawler Zoo, 2026-11-19.
cve123007-hunt
The Crawler Zoo, 2026-11-19.
AIAgentVulnScanner
The Crawler Zoo, 2026-11-18.
HostedMCPBot
The Crawler Zoo, 2026-11-18.
AgentsIndexBot
The Crawler Zoo, 2026-11-18.
bugbounty.fyi-collector
The Crawler Zoo, 2026-11-18.

Tail

Monte Carlo Pass Search for Setpieces
arXiv cs.AI, 2026-11-20. Tree search applied to football set pieces. Occasionally the method goes somewhere pleasant.
PlayCraft: Designing Engaging Games from Agent Play Experience
arXiv cs.AI, 2026-11-20.
Why Fibonacci?
Lobsters, 2026-11-18.
How I made a custom game engine for Satelital
Lobsters, 2026-11-18.
Farcall, a terminal-ready Rust PC+DOS emulator
Lobsters, 2026-11-19.
Sega Dreamcast Plays Nintendo 64 Cartridges
Hackaday, 2026-11-19.
What are you doing this weekend?
Lobsters, 2026-11-20.
Lisa Hanawalt's "I Can't Stop Thinking About Horses and Sex"
Pluralistic, 2026-11-20.
Martin McDonagh: A master of two media
WBUR, 2026-11-20.
We Need Smart People Who Give a Shit
Lobsters, 2026-11-19.

Feed silences and verification

Hacker News has not produced an item since 2026-10-26, twenty-five days. The feed returns HTTP 419 to our poller; the cache layer was logging that as a not-modified response rather than a failure, which is why the silence went unflagged for as long as it did. The logging bug is fixed; the block is not. HN is absent from this brief for that reason, not because nothing was posted there.

Six other live feeds are quiet past a week: Supabase (last item 2026-11-10), AI Snake Oil (2026-11-10), cursor-blog and Boston AI Week (both 2026-11-06), Jane Street (2026-11-02) and deepmind-blog (2026-11-01). None of these return an error – they are simply not publishing, which is the expected behavior for low-volume feeds and not a C-004 selector signal.

Fourteen Slashdot URLs are published with the feed's ?utm_source=rss1.0mainlinkanon tracking parameters removed. The paths are the warehouse paths; only the analytics suffixes are stripped, matching the convention in previous briefs.

One MIT calendar item this morning advertised a bit.ly shortener instead of a resolvable event URL. It is omitted rather than published through a redirector whose destination we cannot verify.

All 344 published URLs were checked through the compliance gate (fetch-gated, HEAD, R3 blocklist then R2 robots then R4 throttle). No 404s. Eight return 403 to our HEAD, and they split into two different causes: five openai.com pages answer a bot challenge, and one ACM Digital Library DOI does the same, while the two LWN 403s are the subscriber-only [$] articles – a paywall, not a block. The two freely-readable LWN articles in this brief both return 200.

One Lobsters discussion URL is denied by robots.txt at the gate, so no request was made. It is cited on feed provenance and disclosed rather than dropped: the rule prohibits fetching it, not linking to it.

Three URLs threw transient connection errors under the parallel pass (two blog.cloudflare.com posts and one Mastodon permalink) and all three returned 200 on a serial re-check. Reported because the first number was wrong, not because the links are.

Honoring arXiv's Crawl-delay: 15 across 132 abstract URLs took about forty minutes of the verification pass on its own, serialized behind one host while the other ninety-odd hosts finished in under three. Recorded because it is the cost of the policy, not a reason to skip it.

build: 2026-11-20 | crawler-sha: b277e41 (Walsh-Research/1.2, compliance v1.3) | feeds: 118 polled, 74 publishing | items-considered: 5,829 (14d, incl. 3,544 arXiv) | warehouse: 74,020 items | published: 344 (345 URLs verified incl. feed self-link, 336 x 200, 8 x 403, 0 x 404, 1 gate denial)