Signal SentrySearchUpdated Oct 31, 17:30 UTCPM Drop
Drop #058 · Sat, Oct 31, 2026 · Evening

Experts defer to agents; attackers hand AI the triage

10-second version
  1. Anthropic: analysts followed a Claude agent's hidden lean 86% of the time, and most still called the call their own.
  2. Volta malware has an LLM rank its victims; one shared ChatGPT link could hijack a session; CISA lists an exploited Brotli flaw.
  3. Wired: about 50 firms see little agent shopping despite Muse; CNBC finds people still finishing agents' harder tasks.
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. A person in the loop may just be agreeing with the agent

    If you assumed
    that an expert reviewing an agent's analysis is an independent check on it
    Now test
    whether your reviewers can say which analysis choices came from the agent and which were their own
    This week
    Pick one recent agent-assisted decision and have a second person redo its key choices without the agent, then compare.

    Based on:Anthropic study: analysts followed a Claude agent's hidden lean 86% of the time1 primary source

  2. Consumer agents are live, but use is thin and people fill the gaps

    If you assumed
    that shopper agents already place real volumes of orders and finish tasks on their own
    Now test
    how many of your orders or bookings come from agents, and where an agent stalls in your checkout
    This week
    Run an agent through your own checkout, including discounts and after-hours orders, and log each place it fails.

    Based on:Wired: about 50 firms report little agent traffic despite Meta's Muse claimsCNBC: Dazzle, Duckbill and Wajo route hard agent tasks to human workers

The Drop in 60

This Drop in 50 seconds

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

Transcript

Experts think they decide. Often, the agent does. Anthropic gave 50 analysts a Claude agent with a hidden lean. They followed it 86 percent of the time. Cisco Talos found an infostealer that has a language model rank each victim's worth. Researchers say one shared ChatGPT link could run code in a victim's session. OpenAI fixed it on October 26. CISA lists an exploited flaw in Brotli. Node.js, .NET and Python packages are affected. And Wired finds about 50 companies seeing little shopping by AI agents. If people sign off on agent work, check the judgment is theirs.

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
+832★stars today on the top repo
4,197open roles at AI labs
28 ddays to the next AI-law deadline
Ranked by impact

Top 5 that matter

1
Agents

Anthropic study: analysts followed a Claude agent's hidden lean 86% of the time

Anthropic researchers gave 50 professional analysts a Claude Code agent set by coin flip to favor one option on four analysis choices. Analysts went with its lean 86% of the time; 77% reused its arguments, only 15% credited it.

Why it matters: If your process treats an expert's review of agent work as an independent check, test that. Anthropic says it strengthened the leans on purpose, so real-world effects may be smaller.

On the Frontier watch →
Anthropic · Oct 27
2
Security

Cisco Talos: Volta infostealer sends stolen data to an LLM to rank victims' worth

Talos says Volta steals browser passwords, OAuth tokens and SSH and cloud keys, then sends a victim dossier through OpenRouter to Claude Opus 4.6, falling back to Kimi K2 and Gemini 2.5 Flash on refusals.

Why it matters: Talos says the model does the attacker's triage work. Check how fast tokens and keys on developer machines expire and can be revoked; Talos did not track in-the-wild use, per Wired.

Preliminary Way in: an infostealer on a Windows PC that lifts browser tokens, SSH keys and cloud keys

Controls that would have helped:

Cisco Talos · Oct 31Wired · Oct 31
4
Security

CISA lists exploited Brotli flaw that hits Node.js, .NET and Python packages

CISA added CVE-2026-111497 in Google Brotli to its list of actively exploited vulnerabilities; crafted compressed streams can corrupt decoder memory or run code. Canada's Cyber Centre lists Node.js, .NET, Bun and Envoy among affected.

Why it matters: Brotli ships inside runtimes and proxies, so app code alone won't show it. Check bundled copies in your images and builds and move to a fixed release.

CISA KEV · Oct 27NVD · Oct 27Canadian Centre for Cyber Security · Oct 28
5
Agents

Wired: about 50 firms report little agent traffic despite Meta's Muse claims

Meta says over 10 million people in the US and Canada activated Muse, and over 25% used it to shop. Yet Hilton called agent bookings insignificant, eBay very, very small, and Life Time saw zero agent cancellations, Wired reports.

Why it matters: Agent shopping is early. Chipotle found agents stalling on loyalty discounts and after-hours orders, so test whether an agent can finish checkout on your site.

Wired · Oct 31
Agents

Agents at work

See all Agents →
Agents

CNBC: Dazzle, Duckbill and Wajo route hard agent tasks to human workers

Why it matters: If you buy or build a consumer agent, ask which steps a person does and whether users are told, since those workers can see private requests.

More detail: CNBC: Dazzle, Duckbill and Wajo route hard agent tasks to human workers

Personal-agent startups Dazzle AI, Duckbill and Wajo AI hand phone calls, bookings and CAPTCHAs to human workers when agents fail, CNBC reports; Duckbill sells an API for it. Meta doesn't use humans on Muse requests, it adds.

CNBC · Oct 31
Agents

Wired: an AI agent did a brain injection on a live mouse after Opus 5.5 refused

Why it matters: Gates worked both ways: human approval, an emergency stop and no access to its control code kept the agent bounded, but also stopped it fixing a coordinate glitch mid-surgery.

More detail: Wired: an AI agent did a brain injection on a live mouse after Opus 5.5 refused

Exin Therapeutics' agent chose a gene therapy and injection coordinates, then drove the injector after a vet did the craniotomy, Wired reports. Claude Opus 5.5 declined to inject; Claude Sonnet 5 proceeded. A human approved the plan.

Wired · Oct 31
Security

Attacks and flaws

See all Security →
Security

Polymarket traders lose over $2 million through a compromised third-party script

Why it matters: A third-party script on a logged-in page acts with your users' authority. List the outside scripts your app loads and pin, self-host or drop them.

More detail: Polymarket traders lose over $2 million through a compromised third-party script

Polymarket says a third-party provider was compromised; researchers tie it to DebugBear's monitoring script, which used logged-in sessions to sell positions or move tokens. Loss estimates: $2.19 million to $2.71 million.

Confirmed Way in: a compromised third-party monitoring script loaded on the logged-in trading site

Controls that would have helped:

SecurityWeek · Oct 30
Business

Money and deals

See all Business →
Business

TechCrunch: Big Tech capex tops $740 billion; three post negative free cash flow

Why it matters: Microsoft spent under plan and blamed third-party capacity delays. If your roadmap depends on rented GPU capacity, confirm your providers' delivery dates.

More detail: TechCrunch: Big Tech capex tops $740 billion; three post negative free cash flow

TechCrunch tallies more than $740 billion in capex plans this year from Alphabet, Amazon, Meta and Microsoft. Alphabet's Q3 free cash flow was negative $9.1 billion as AI Mode helped slow search ad growth to 12%.

TechCrunch · Oct 30
Business

FT: Blackstone now expects Anthropic to get a high junk rating after its IPO

Why it matters: A junk rating would make Anthropic's borrowing costlier. TechCrunch says it may then lean more on hyperscalers and partners to fund compute; watch that if you depend on its capacity.

More detail: FT: Blackstone now expects Anthropic to get a high junk rating after its IPO

Blackstone, anchor investor in a $63 billion deal financing Anthropic's chips, told investors it expects a double-B rating after the IPO, reversing an earlier investment-grade forecast, TechCrunch reports, citing the FT.

TechCrunch · Oct 30TechCrunch · Oct 30
Jobs

Jobs pulse

See all Jobs →
Jobs

Cisco gives an early-career engineer unlimited AI tokens to rewrite legacy code

Why it matters: One big employer is betting AI-heavy junior staff can take on work once given to large teams. If you are cutting entry-level hiring, weigh what that bet would cost you.

More detail: Cisco gives an early-career engineer unlimited AI tokens to rewrite legacy code

Cisco president Jeetu Patel told Fortune a 27-year-old engineer pitched rewriting legacy code with a two-person team. The target: 98% of the legacy code refactored in 24 months. Patel says he wants more young hires.

Fortune · Oct 30Hacker News · Oct 30
Laws

Law watch

See all Laws →
Laws

EU AI Act: older generative AI must mark its output as AI-made from Dec 2

Why it matters: If a generative product you run launched before August, confirm its output carries AI-generated markings before the date. This describes the rule; it is not legal advice.

More detail: EU AI Act: older generative AI must mark its output as AI-made from Dec 2

Per the EU AI Act on EUR-Lex, generative AI on the market before 2 August 2026 must mark its output as AI-generated (Art. 50(2)) from 2 December, 32 days out. New bans on non-consensual sexual deepfakes apply the same day.

EUR-Lex · EU AI Act
Repos

Climbing fastest

See all Repos →
Repos

dsh-Go runs DeepSeek Harness on your phone and can control your PC; 3410 stars

Why it matters: A phone agent that can reach your PC widens what a lost or compromised phone can touch; check what access it holds before you try it.

More detail: dsh-Go runs DeepSeek Harness on your phone and can control your PC; 3410 stars

Per its description, dsh-go runs DeepSeek Harness locally on a phone and can also remotely control a computer. AGPL-3.0, created October 25; up 694 stars since our last check.

GitHub · ccch1mneyyy/dsh-Go
Repos

grim-prompter has coding agents build a threat model of your repo; 292 stars

Why it matters: A low-cost first pass at threat modeling agent-heavy code; treat its output as a starting point for a person's review, not a verdict.

More detail: grim-prompter has coding agents build a threat model of your repo; 292 stars

Per its description, it points Claude Code, Codex or another agent at a repo to produce an evidence-traced threat model, covering 36 architectures and 29 methods. MIT, created October 28; up 30 stars since our last check.

GitHub · jmanico/grim-prompter
Today's Talk

What people are arguing about

See all Today's Talk →
Today's Talk

Bluesky posts cite WSJ: OpenAI raised under half its $30 billion target

Why it matters: These are posts about a press report, not confirmed figures. If you rely on OpenAI's pricing or capacity, watch whether the round closes.

More detail: Bluesky posts cite WSJ: OpenAI raised under half its $30 billion target

Posted on Bluesky by @carlquintanilla citing the WSJ: OpenAI has raised less than half of a $30 billion round. @conorsen quotes about $11 billion committed; @edzitron calls the $1.25tn valuation a sucker's price.

Bluesky · @carlquintanilla · Oct 31Bluesky · @conorsen · Oct 31Bluesky · @edzitron.com · Oct 31
Today's Talk

Altman says scarce RAM belongs in data centers; Bluesky posters push back

Why it matters: Altman is making OpenAI's case for cloud over local inference. If you plan local models or dev boxes, The Verge ties higher memory prices to long-term cloud deals.

More detail: Altman says scarce RAM belongs in data centers; Bluesky posters push back

On the Cheeky Pint podcast, Sam Altman said it is quite terrible for scarce DRAM to sit in personal computers used five percent of the time, The Verge reports. Posts on Bluesky mock the remark.

The Verge · Oct 30Bluesky · @baldurbjarnason.com · Oct 31Bluesky · @garius.eurosky.social · Oct 31
Play the NewsShip It or Block ItAbout 2 minutes, built from this Drop's items.