Morning Brief: Sunday, November 22

Seventy-four feeds reporting. Two weeks. 5,523 items reduced to what follows. (what we track, how we crawl, subscribe)

Sunday's thread is the plumbing between agents. The week's hardest evidence is not about model capability – it is about what happens to data and identity once a request crosses from one agent service into the next.

Two million canaries is the number. Dan Goodin's 2026-11-20 report describes an experiment that planted unique, trackable secrets inside prompts and followed them through a network of AI services; the secrets did not stay put. That is a measurement, not a prediction, and it lands in the same days that Signal began restricting mobile screenshots – explicitly framing agents reading the screen as an existential threat rather than a feature – and that arXiv's cs.AI stream carried four separate papers on the same seam: tool names that are not identities, agent cards at the A2A trust boundary, zero-trust propagation across delegation chains, and authorization that can actually be audited across the model, harness, and workflow layers. Our own crawler census agrees from the other direction: of twenty-seven new bots recorded in the zoo over two days, a visible share exist to scan MCP endpoints, inventory tools, read llms.txt, or monitor other agents.

The second thread is narrower and older: if agents write the code, what is the artifact a human reads? The New Stack asked it directly. Nvidia's answer was 20,000 training examples aimed at its own model's mistakes. Graham Dumpleton's answer, visible in Simon Willison's link-blog, is a daily public development log for an AI-assisted Rust server – a post nearly every day for three weeks, the narrative shipped alongside the binary. Martin Fowler's Deterministic Simulation Testing is a third answer, and arXiv's contribution is the deflating finding that for automated code review, a single agent is often enough.

Top (5-7 min)

Researcher plants 2m canaries into the AI supply chain. A privacy disaster surfaced.
Ars Technica via Lobsters, 2026-11-21. Dan Goodin, dated 2026-11-20. Unique trackable secrets embedded in prompts, followed through a chain of AI services, leaked. The method is the contribution: tripwires, not surveys.
Signal restricts mobile screenshots as it confronts 'existential threat' of AI surveillance
TechCrunch, 2026-11-21. A messaging app treating on-device agents as adversaries. The screen, not the network, is now the exfiltration surface.
When Tool Names Are Not Identities: Measuring Provider Misbinding in MCP-Based AI Agents
arXiv cs.AI, 2026-11-17. Measures the gap between a tool's advertised name and the provider actually serving it. The naming layer is load-bearing and has no authentication.
If we stop reading the code, what do we read?
The New Stack, 2026-11-21. The question the rest of the week keeps answering indirectly. Worth reading before the papers below, which are all partial answers.
Mapping the AI Supply Chain: Dependencies, Vulnerabilities, and Market Concentration in the AI Ecosystem
arXiv cs.AI, 2026-11-19. The structural companion to the canary experiment: where the dependencies concentrate, and therefore where a leak propagates furthest.
Nvidia's Accounting Practices Now Being Investigated by America's SEC
Slashdot, 2026-11-21. Arrives the day after California's AG subpoena over chip sales to China. Two separate investigations, one week.
Auditable Authorization for LLM Agents: Empirical Evidence Across Model, Harness, and Workflow Layers
arXiv cs.AI, 2026-11-17. Separates authorization failures by layer, which is the prerequisite for assigning them to an owner.

Themes

Two million canaries

The canary experiment and the MCP identity papers are the same finding approached from opposite ends. Goodin's report is empirical: plant secrets, watch them surface somewhere they should not be. The papers are structural: When Tool Names Are Not Identities shows that the string an agent uses to select a tool does not bind to a provider, We Have All the Cards locates the same problem at the Agent2Agent protocol's operational trust boundary, and Agent-ZTA proposes propagating security state bidirectionally across a delegation chain instead of trusting it at the first hop.

We Have All the Cards: Securing Agent Cards at the Operational Trust Boundary of the Agent2Agent Protocol
arXiv cs.AI, 2026-11-17.
Agent-ZTA: A chain-Aware Zero Trust Model for Bidirectional Security-State Propagation and Runtime Containment in Multi-Agent Delegation
arXiv cs.AI, 2026-11-17.
Can You Trust Your Finance Agent? Drift, Diagnosis and Auditable Intervention on One Shipped Financial Research Agent
arXiv cs.AI, 2026-11-16. One shipped agent, instrumented. The shipped qualifier is the reason to read it.
Readie: Low-Latency Code Sandboxes for AI Agents
arXiv cs.AI, 2026-11-18. Containment at the execution layer rather than the protocol layer.

What you read when nobody reads the code

Four answers landed in two days, none of them the code. Nvidia's is data: 20,000 examples targeting its own model's errors, with a coding score to show for it. Dumpleton's is narrative: a development log published as the work happens, running near-daily since 2026-11-04, which Willison collected. Fowler's is a technique – Deterministic Simulation Testing, which replaces reading with replayable execution. And arXiv's is a negative result: for automated code review, the multi-agent architecture is often unnecessary.

Nvidia used 20,000 training examples to fix its AI's mistakes. Its coding score jumped.
The New Stack, 2026-11-21.
Muru
Simon Willison, 2026-11-21. Graham Dumpleton's AI-assisted Rust server for WSGI and ASGI Python apps, twenty years after he started mod_wsgi. Willison collects the running development log; the cadence is the point as much as the server is.
Deterministic Simulation Testing
Martin Fowler, 2026-11-17.
When Single Agent Is Enough: Rethinking Multi-Agent Systems for Automated Code Review
arXiv cs.AI, 2026-11-19.
PrivacyPR: A Dataset of Real-World Privacy Fixes for Benchmarking Privacy-Aware Code Review
arXiv cs.AI, 2026-11-16. Real privacy fixes as ground truth, which makes the benchmark harder to game than synthetic defects.

Nvidia's two stories

The SEC accounting investigation and the California subpoena are separate matters landing in the same week, and TechCrunch's framing treats the company as two stories running in parallel – the engineering organization and the governance one – following Jensen Huang's exit.

A tale of two Nvidias
TechCrunch, 2026-11-21.
Bitcoin Drops to $56,085, a Two-Year Low, in Worst Month Since 2022
Slashdot, 2026-11-21. The market backdrop, included because the Nasdaq erased its 2026 gains in the same stretch.

Scan

Proof automation and formal methods

Last week's Riemann claims have not been settled, and this week's papers are the infrastructure question underneath them: what does a machine-checkable proof pipeline actually look like.

Even more posts, opinion pieces, interviews, and position statements
Terence Tao, 2026-11-22.
Hurried Announcement: Alessandro Masullo GRH
Math 216, 2026-11-21. A seminar announcement on the Generalized Riemann Hypothesis, posted at short notice.
Turing Agent: Agentic Formal Verification for Mathematical Research
arXiv cs.AI, 2026-11-17.
From Sorry to QED: Refining LLM Math Proofs via Angelic Execution
arXiv cs.AI, 2026-11-18. The title names the failure mode: Lean's sorry as the gap between a plausible proof and a checked one.
Reusing formal Vlasov-Maxwell proofs to certify plasma solver outputs
arXiv cs.AI, 2026-11-19.
From Pen-and-Paper to Machine-Checked Proofs: Lean for System Identification
arXiv cs.AI, 2026-11-16.

Security and the attack surface

GhostScript CVE 2026-125396: "Upload2Shell"
Lobsters, 2026-11-21.
Microsoft 365 Outage - Post Incident Report
Lobsters, 2026-11-21. The 2026-11-12 outage write-up, collected outside Microsoft's own status channel.
Trump says reported FBI database hack is 'not good'
TechCrunch, 2026-11-21.
US DOJ Charges Suspected Russian Agent for Stalking Palantir Employees in California
Slashdot, 2026-11-21.
Will AI-Fueled Cyberattacks Lead to Better Funding for Critical Open Source Projects?
Slashdot, 2026-11-21.
port scanning for mojang
Lobsters, 2026-11-21.
Network-Exposed Configuration Attacks on LLM-Controlled Social Robots
arXiv cs.AI, 2026-11-18. The embodied version of the same trust-boundary problem: the configuration endpoint is on the network.

Policy, society, and the human read

The Claude Delusion
Lobsters, 2026-11-22. Jan Schaumann, dated 2026-11-13, surfaced on Lobsters today. An essay on wealth, illusory superiority, and Dunning-Kruger in tech leadership rather than on any model's behavior.
Internet Archive Staffers Are Unionizing
Slashdot, 2026-11-21.
Is selling an AI version of your face to this microdrama startup the future of Hollywood?
TechCrunch, 2026-11-21.
Alaska, your next governor is a Wikipedia nerd
TechCrunch, 2026-11-21.
Americans Enters Post-Penny World, Businesses Now Allowed to Round Cash Transactions to Nearest Nickel
Slashdot, 2026-11-21.
Wrong with Confidence: Measuring Hallucination, Overconfidence and Identity-Conditioned Behaviour in Large Language Models
arXiv cs.AI, 2026-11-17.

The crawler zoo

Thirty-four crawler-zoo arrivals are dated 2026-11-21 or 2026-11-22; twenty-seven are new to this series, and twenty-six of those are listed here. A visible share of them exist to inspect agent infrastructure rather than content: MCP tool scanners, a tool inventory checker, an llms.txt scanner, and an agent monitor.

MCP and agent infrastructure
secdb-mcptoolscan, mcp-inventory-check, mocop-crawler, LLMS-Txt-Scanner, ZeroFox-AI-Agent-Monitor, free-apis-check
Commerce and catalog
catalog-indexer, catalog-checker, LocalEcommerceAssetResearch
Outreach and contact harvesting
outreach-scout, outreach-poc, ContactXBot, bottomlines-crawl
Audit, archive, and search
AuthorizedPublicPageAuditor, PubSleuthBot, bxzxxr-archive, watrsearch, PageHubBot
Unclassified
Chrome-Lighthouse, keda, e5-frame-b, AterstodBot, NoahBot, Jambot, ProngBot, WallingBot

Tail

What is a Scamp?
udamonic, 2026-11-21. A small dedicated Forth computer. Filed here because it is the opposite of everything above: one language, one board, no network.
Scientists Discover New and Horrible Eagle-Based Way Our Ancestors Died
404 Media, 2026-11-21.
How 2 Security Research (MLG)
Lobsters, 2026-11-21.
Decapitation Seems Like a Great Way to Start X-Mas Decorating
Hackaday, 2026-11-21.

Feed silences and verification

118 feed endpoints were polled this morning. 46 returned new content, 71 returned 304 Not Modified, and one errored.

www.bornholt.dev failed DNS resolution again. The feed has never resolved in any recorded poll, so this is a dead endpoint rather than an outage, and it is counted as such rather than as a silence.

Hacker News has recorded nothing since 2026-10-26. This is our bug, not a feed silence: the endpoint answers 419 and our cache layer records it as :not-modified, which makes it invisible in the health report. It is unchanged from yesterday and still unfiled – bd is not on PATH in this environment.

No arXiv announcement landed on 2026-11-21 or 2026-11-22. The cs.AI feed's newest items are dated 2026-11-20, and the same two-day gap appears at 2026-11-14/15 and 2026-11-21: it is the announcement calendar, not a selector break. Because the 14-day arXiv pool is therefore identical to yesterday's, today's eighteen papers were selected against this brief's threads – agent identity, code review, proof automation – rather than by recency, and all eighteen are new to this series.

Thirteen 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 of the twenty-seven new crawler-zoo arrivals carries a racial slur as its user-agent string. It is recorded in the warehouse and omitted from the list above; the count of twenty-seven includes it so the census stays accurate.

All 102 published URLs were HEAD-checked through the compliance gate (fetch-gated, R3 blocklist then R2 robots then R4 throttle), parallel across hosts and serial within each host. All 102 return 200 – no gate denials, no 403s, no 404s, no connection errors. The LWN subscriber-only 403s that usually appear in this section are absent because no [$] article is cited today.

Honoring arXiv's Crawl-delay: 15 across the eighteen abstract URLs costs four and a half minutes on its own – serialized behind one host, while the other thirty-seven hosts finish in parallel. Recorded because it is the cost of the policy, not a reason to skip it.

build: 2026-11-22 | crawler-sha: b277e41 (Walsh-Research/1.2, compliance v1.3) | feeds: 118 polled, 74 publishing | items-considered: 5,523 (14d, incl. 3,272 arXiv) | warehouse: 74,267 items | published: 102 (102 URLs verified through the gate, 102 x 200, 0 x 403, 0 x 404, 0 gate denials, 0 errors)