Signal SentrySearchUpdated Nov 1, 17:30 UTCPM Drop
Drop #060 · Sun, Nov 1, 2026 · Evening

Sites start shutting the door on AI agents

10-second version
  1. CNBC: JPMorgan, Amazon and other big brands block personal agents; an audit finds most retailers do it by accident.
  2. Binance resumes withdrawals after a $6.3 billion hack the FBI ties to North Korea; CISA flags exploited PHP flaws.
  3. Three preprints find agents' task scores overstate how well they serve users, follow procedures or finish work.
Our read

What changed

Every signal →

Our reading of this Drop's sources: what to rethink, and one thing worth checking this week. Not a prediction.

  1. Passing the task is not the same as serving the user

    If you assumed
    that an agent's task-completion score shows how well it will serve your users
    Now test
    whether your agent evals check users' own phrasing, required steps, unrequested changes and fully finished tasks, not just end state
    This week
    Add one check to your agent evals for skipped required steps or unrequested changes, and score fully finished tasks separately

    Based on:Preprint: computer-use agents' task completion falls 36% when personas phrase requestsPreprint: outcome-only checks wrongly pass 6.5-12.0% of banking-agent sessionsPreprint: best coding agent meets 69.9% of criteria but fully finishes 26.4% of tasks3 primary sources

  2. Sites now decide whether customers' agents get in

    If you assumed
    that a customer's permission is enough for their agent to act on a bank or store site
    Now test
    whether the sites your agent relies on block agents on purpose or through old bot rules
    This week
    List the sites your agent touches and test each one against its published access rules

    Based on:Banks and retailers are blocking personal agents like Muse, CNBC reports

The Drop in 60

This Drop in 53 seconds

Tap to play. It never plays by itself, and the captions are in the picture.

Transcript

Banks and stores are shutting out AI agents. CNBC reports JPMorgan, Schwab, Amazon and Kohl's block personal agents. One audit found 37 of 51 big retailers block shopping agents, mostly by accident. Binance reopened withdrawals after a 6.3 billion dollar hack the FBI pins on North Korea. How attackers got in is still unclear. A preprint finds computer-use agents complete 36 percent less when simulated users phrase requests their own way. CISA lists two PHP flaws as actively exploited. Check your versions. If your agent acts for users, test it against each site's rules before your customers do.

Made with AI from this Drop's sourced items. The facts come from the sources linked on each item. Script: Claude · Voice: OpenAI gpt-4o-mini-tts · Visuals: motion graphics

425models priced on OpenRouter
+683★stars today on the top repo
4,196open roles at AI labs
27 ddays to the next AI-law deadline
Ranked by impact

Top 5 that matter

1
Agents

Banks and retailers are blocking personal agents like Muse, CNBC reports

CNBC reports JPMorgan, Schwab, Amazon, Neiman Marcus and Kohl's block personal agents like Meta Muse. An audit of 51 big U.S. retailers by Billions Network found 37 block shopping agents mostly by accident, via old bot rules.

Why it matters: If your product sends agents to shop or bank for users, check target sites' bot rules first: many block agents by default, and none in the audit lets a verified agent check out.

CNBC · Nov 1
2
Security

Binance resumes withdrawals after a $6.3 billion hack; FBI blames North Korea

Binance says roughly 90% of stolen user assets are now backed again, CNBC reports. Binance first pointed to a compromised DevOps engineer's laptop; its co-CEO later said the attack was more complex. How it happened is still unclear.

Why it matters: Even with the cause unconfirmed, the first suspect was one engineer's laptop: check which developer machines can reach systems that move money or sign transactions.

CNBC · Nov 1
3
Agents

Preprint: computer-use agents' task completion falls 36% when personas phrase requests

Researchers report in a preprint: 74 fixed tasks, 44 simulated personas, nine backbones, 3,330 runs. Fully completed persona requests won top satisfaction in only 62.7% of cells, vs 84.1% for original requests.

Why it matters: A benchmark pass on a clean instruction may overstate how your agent serves real users; test with users' own phrasing, missing details and requests to stop.

On the Frontier watch →
Hugging Face Papers · Oct 16
4
Security

CISA lists exploited PHP and Linux kernel flaws that can be chained into webshells

CISA added two PHP flaws, CVE-2026-109618 and CVE-2026-109619, each allowing unauthenticated remote code execution, to its actively exploited list. Kernel flaw CVE-2026-46300 can be chained with them to plant PHP webshells.

Why it matters: If you run PHP apps or self-host on Linux, check your PHP and kernel versions against vendor fixes this week.

CISA KEV catalog · Oct 28NVD · CVE-2026-109618NVD · CVE-2026-109619NVD · CVE-2026-46300
5
Business

CNBC: Santander says AI generated 194 million euros in value in nine months

CNBC reports Santander credits AI with 194 million euros of business value in nine months of 2026. Novartis and Standard Chartered say AI is reaching their results; Carlsberg says gains aren't at scale yet.

Why it matters: Buyers can now ask vendors for what big firms report: a measured AI contribution to revenue or cost, not just usage figures.

CNBC · Nov 1
Agents

Agents at work

See all Agents →
Agents

Preprint: outcome-only checks wrongly pass 6.5-12.0% of banking-agent sessions

Why it matters: If your agent evals only check end state, add checks for required steps and unrequested changes, especially where a rule demands a step first.

More detail: Preprint: outcome-only checks wrongly pass 6.5-12.0% of banking-agent sessions

TauIndianBankBench has 1,000 Indian retail-banking tasks and checks every required step, not just final records. For every model, 6.5-12.0% of sessions looked successful on records alone despite a skipped step or unrequested change.

Hugging Face Papers · Sep 30
Agents

Preprint: best coding agent meets 69.9% of criteria but fully finishes 26.4% of tasks

Why it matters: Partial credit hides unfinished work: when you grade coding agents, count tasks fully done, and expect more misses on edits that span several kinds of change.

More detail: Preprint: best coding agent meets 69.9% of criteria but fully finishes 26.4% of tasks

Co-Design Bench draws 273 tasks from 26,314 real user coding sessions. Under one shared harness, the best of ten model endpoints passed 69.9% of criteria per task on average, but fully satisfied only 26.4% of tasks.

Hugging Face Papers · Oct 26
Agents

Tom's Hardware: 50-agent Minecraft colony stalled in bureaucracy; GPT beat Claude 6-1

Why it matters: More agents can mean more coordination, not more output; before scaling a swarm, measure whether more agents finish more work than fewer.

More detail: Tom's Hardware: 50-agent Minecraft colony stalled in bureaucracy; GPT beat Claude 6-1

Researchers ran 50 GPT-6.1 Sol agents to build a colony in three hours; it never thrived and drifted into bureaucracy, Tom's Hardware reports. In Mindustry, GPT agents beat Claude Opus 5.5 agents 6-1 over 7 games.

On the Frontier watch →
Tom's Hardware · Nov 1Hacker News · Nov 1
Business

Money and deals

See all Business →
Business

CNBC: Meta's ad revenue grew 26% while Google Search grew 12% as AI shifts habits

Why it matters: If you buy search ads or depend on search traffic, check how much discovery now happens in chatbots; analysts quoted by CNBC see search revenue under pressure.

More detail: CNBC: Meta's ad revenue grew 26% while Google Search grew 12% as AI shifts habits

Meta's ad revenue rose 26% to $63.1 billion; Google Search grew 12% to $63.4 billion, CNBC reports. Meta's CFO says advertisers following its AI assistant's recommendations saw almost 5% more conversions.

CNBC · Nov 1
Business

Google: Oracle will be first hyperscaler to offer TPUs outside Google Cloud, in 2027

Why it matters: TPU capacity may become available outside Google's own cloud, so teams planning 2027 inference can weigh more chip options across providers.

More detail: Google: Oracle will be first hyperscaler to offer TPUs outside Google Cloud, in 2027

Google's October infrastructure roundup says its TPUs arrive on Oracle Cloud in 2027. It also lists Agent Compute Engine, microVM sandboxes for agents, and llm-d 1.0, an open-source Kubernetes serving stack.

Google Cloud · Oct 30
Repos

Climbing fastest

See all Repos →
Repos

kanzo runs a self-hosted kanban board for humans and agents; 422 stars

Why it matters: If you run several coding agents, a shared board is one way to see their work next to people's; check the AGPL license before embedding it.

More detail: kanzo runs a self-hosted kanban board for humans and agents; 422 stars

kanzo, AGPL-3.0 and created Oct 25, describes itself as one 4 MB binary that starts in 2 ms with zero setup. It gained 195 stars since our last check.

GitHub · emanuele-em/kanzo
ReposFollowing since Oct 30, 2026in 4 Drops

Mika runs an on-device Android agent that operates apps and the web; 3574 stars

Why it matters: Phone agents that drive apps directly are drawing fast interest; if your app has a mobile front end, check how it behaves when an agent operates it.

More detail: Mika runs an on-device Android agent that operates apps and the web; 3574 stars

Mika, Apache-2.0 and created Oct 29, is an on-device Android agent that operates apps and the web. It gained 683 stars since our last check, the most of any repo we track this run.

GitHub · wangzan101/Mika
Today's Talk

What people are arguing about

See all Today's Talk →
Today's Talk

Benchmark of agent-built apps by JS framework draws 182 points on Hacker News

Why it matters: Claims that a framework suits agents better can be tested cheaply; run your own repeated builds before rewriting a frontend for agents.

More detail: Benchmark of agent-built apps by JS framework draws 182 points on Hacker News

Nolan Lawson had Opus 5.5 build the same Pokémon app in frameworks from Next.js to Hono+htmx, 3-5 runs each. He writes SvelteKit, Gracile+Lit and Angular cost more to build, and Angular failed to produce a working app twice.

Nolan Lawson · Oct 28Hacker News · Nov 1
Today's Talk

Posts say DeepSeek open-sourced a Super-Agent coding prototype

Why it matters: None of this is confirmed by DeepSeek in our sources; if you weigh open coding agents, read DeepSeek's own release before testing it on real repos.

More detail: Posts say DeepSeek open-sourced a Super-Agent coding prototype

Posted by @Polymarket and @Cointelegraph: DeepSeek released Super-Agent, an open-source coding agent prototype pitched against Claude Code. @kimmonismus posted that a router session hands work to parallel workers.

X · @Polymarket · Nov 1X · @Cointelegraph · Nov 1X · @kimmonismus · Nov 1
Today's TalkFollowing since the Morning Dropin 2 Drops

Posts say Anthropic is adding invisible text watermarks to more Claude models

Why it matters: If the posts are right, text your product generates with these models may carry a hidden mark; check vendor docs before promising customers anything about provenance.

More detail: Posts say Anthropic is adding invisible text watermarks to more Claude models

@alex_prompter posted that Anthropic will extend its watermark to Sonnet 4.6, Opus 4.6, Opus 4.7 and Mythos 5, with Opus 5.5 already marked. @EricDJuly posted that Grok did the same, citing the EU AI Act.

X · @alex_prompter · Nov 1X · @EricDJuly · Nov 1
Play the NewsAI ConnectionsAbout 3 minutes, built from this Drop's items.