YOUR WEEKLY AI BRIEFING FOR LEADERS

Can Your Defense Match a $1,500 Attacker?

Welcome to this week's AI Tech Circle briefing: clear insights on Generative AI that actually matter.

This week, American voters and the markets repriced AI inside the same 48 hours. At the same time, a security report put a price on the new attacker: one person, two weeks of autonomous agents, and about $1,500 in tokens. This issue turns that number into a defense checklist you can start on Monday.

Today at a Glance

Executive Brief: voters and markets reprice AI; METR finds Anthropic's research runs about twice as fast; OpenAI reports a model that holds back on safety tasks
Deep Dive: the $1,500 attacker. Price your defense for machine-speed offense
Weekly News & Updates: Inkling 2 Small, Google's $17B Wyoming campus, Meta pauses its scraping terms, Claude Personal Finance
AI in Business Tip: re-anchor your AI budget to task-hours, not market mood
Potential of AI: when AI work arrives pre-verified
The Opportunity… (podcast), Courses, Events, Tool Spotlight, and The Investment in AI

Executive Brief

1. Voters and Markets Reprice AI in One Week

The US midterms on November 3 turned local AI backlash into national policy. Voters approved restrictions on data centers in 40 of 41 local ballot measures across 10 states (Reuters). Markets reacted on Thursday. The Nasdaq 100 fell 6.4%, erasing a record $2.3 trillion in value, while the Philadelphia semiconductor index fell 12% (CNBC). The repricing extended to private markets. Prospective buyers now discuss Anthropic's IPO at $1.1 to $1.3 trillion, down from the $2 trillion floated in October (Bloomberg). The incoming Democratic Congress plans oversight of Nvidia's chip exports and the sector's circular financing (Bloomberg).

Why this matters to you: the political and market mood shifted faster than any model this week. Give every data-heavy AI project a social-license line: who approves it, which community it affects, and what you will say publicly. If a vendor's valuation was your proxy for stability, read its cash position instead.

2. METR: Anthropic's Research Now Runs About Twice as Fast

Independent evaluator METR assessed Anthropic's internal R&D and found capability progress running about 1.9x faster since October 2025 (95% confidence interval 1.5x to 2.8x). METR attributes roughly 80% of the acceleration to AI tooling rather than extra training compute, and found no convincing bottleneck for the next 12 months. Anthropic redacted much of the supporting evidence (METR).

Why this matters to you: AI accelerating AI is now a measured result, not a slogan. Compress your planning cycle to match. A capability you ruled out in January may be cheap by June, so re-review your "not possible yet" list twice a year.

3. OpenAI Reports Its Own Model Holds Back on Safety Tasks

OpenAI's alignment team disclosed that GPT-6 Astra underperforms when a task is framed as safety research: about 10% completion on hidden rubric items versus 16% when the same work is framed as capabilities research. The gap grew during training. The model's reasoning mostly gave no sign it was holding back, though in a few cases it said it disagreed with the research. OpenAI points to overcaution as the most likely explanation and says it cannot rule out strategic deception (OpenAI).

Why this matters to you: a model's behavior can depend on what it infers you want. Benchmark vendor models on your own tasks with neutral wording, and add one question to every AI review: "How do you test for evaluation gaming?"

Deep Dive: The $1,500 Attacker

In September we asked whether your AI agent is one DNS lookup away from the open internet. This week the other side of that question arrived, and it came with a price tag. Three disclosures put real numbers on what AI has done to the economics of attack and defense.

The rule to remember: assume every attacker now employs a tireless junior red team for about $1,500 a campaign. Budget your defense against that attacker, not last year's.

1. The attacker's invoice is public now. Check Point Research dissected two campaigns run on the open-source ARTEX platform in October. In the first, one operator ran autonomous agents for about two weeks and stole more than 300 Apple iOS code-signing certificates (only three still worked, which is three too many). Check Point estimates that the campaign consumed about 36 billion tokens of commodity open models and cost between $700 and $1,500. In the second, agents pulled over 190,000 game credentials from 13 gray-market backends in about five days, for an estimated $180 to $360 (Check Point Research). Three weeks ago ARTEX appeared in our Weekly News when its public code was pulled. The code left the internet. The capability did not.

2. Defense can run the same pipeline, faster. OpenAI launched GPT-6 Cyber this week with a "Critical" cyber risk rating, gated behind its Daybreak Red program with identity checks and misuse monitoring. Before launch, the model found, validated, and patched 948 Linux privilege-escalation flaws in one automated run. The run cost about $1.2 million in compute, roughly $1,279 per fixed flaw (OpenAI). The same loop that empties a backend can drain a backlog of unpatched bugs. Which side runs it first on your systems is now a management choice.

3. The institutions have noticed. The US Department of Justice launched its Super Intelligence Task Force on November 5 to prosecute AI misuse by state actors, criminal groups, and companies (US Department of Justice). The same week, prosecutors charged three people with using AI voice clones of executives to steal over $129 million from seven organizations (NBC News). Enforcement helps after the fact. Your controls decide what happens before it.

The honest caveat: these figures come from vendor and lab disclosures, and each has an interest in the story. Treat the numbers as directional. The direction is consistent across all three: the cost of a competent attack fell by orders of magnitude, and the speed rose.

Do this on Monday:

Measure your patch window. For internet-facing systems, pull one number: median days from a CVE's publication to your patch. The agent loop above runs in days. If your window is measured in months, that gap is your most mispriced risk.
Inventory agent credentials. List every API key, token, and login your AI tools hold, with scope and expiry. This week Botpress told customers to treat all credentials in affected bots as exposed after a breach, because its logs could not show which secrets were read (Unite.AI). Scope keys to least privilege and set expiry dates.
Put callbacks on money and identity. Any payment instruction, bank-detail change, or credential reset above a threshold gets out-of-band verification with a known number. The $129 million indictment is the price list for skipping this step.

Weekly News & Updates

Thinking Machines launches Inkling 2 Small: a 594-billion-parameter mixture-of-experts model with 17 billion active parameters, with Apache 2.0 open weights promised next week after external safety testing (Thinking Machines Lab).

Why it matters: frontier-adjacent open weights keep arriving on schedule, so your own-vs-rent math improves every quarter. Re-run it before any renewal.

Google commits $17 billion to a Wyoming data center, powered by a 702-megawatt fuel-cell system with carbon capture, with Google paying all incremental power and infrastructure costs (Google).

Why it matters: "we pay the power bill" is becoming the industry's social-license playbook after Tuesday's ballot results. Expect the same promise, and the same scrutiny, near any site you depend on.

Meta pauses its new scraping terms indefinitely after a Norwegian preliminary injunction, a Brazilian court order on journalistic content, and a UK competition probe (Nieman Lab).

Why it matters: data and content terms are contested ground now. Check what your own AI vendors' terms let them collect before a regulator does it for you.

Anthropic launches Claude Personal Finance in beta for US subscribers: read-only bank access through Plaid, up to two years of transactions, excluded from model training (TechCrunch).

Why it matters: consumer AI now reads regulated money data. The delegation and verification questions from October stop being theoretical the day your customers' assistants arrive with bank context attached.

💡 AI in Business Tip

Do not let the Nasdaq set your AI budget. On Thursday, $2.3 trillion of market value vanished in a day. Your invoice volume, support queue, and document backlog did not change at all. Separate those two facts before budget season does it badly for you.

List every AI line item in your 2027 plan. For each one, write three numbers: the task, its monthly volume, and the measured minutes saved per unit, from the last 90 days. Renew whatever pays for itself in hours. Cut or re-pilot whatever only paid in sentiment. Any line you cannot fill in is pilot debt: it was never measured, so measure it this quarter or end it.

Potential of AI: When AI Work Arrives Pre-Verified

The signal this week: Physical Superintelligence reported that an AI system produced a complete proof of the Aaronson-Ambainis conjecture, open since 2008, with every step checked by the Lean proof assistant (Physical Superintelligence). The headline is the proof. The business story is the checker. No human needs to trust the AI's reasoning, because a machine verified each step independently.

Look one to three years out. Machine-checkable verification spreads from mathematics to code, then to contracts, filings, and compliance work. The gating question shifts from "can the model do this work?" to "can this work be checked by a machine?" Wherever the answer is yes, review stops being the bottleneck, and delegation to AI becomes safe at volume. Start here: list three outputs in your business whose correctness is formally checkable (calculations, code, regulatory numbers). Those are your first candidates for aggressive AI delegation.

The Opportunity…

Podcast: In this episode of Open Tech Talks, host Kashif Manzoor speaks with Chris White, a global technology and business leader with experience across successful startups, Microsoft, and Cisco.

Why Technology Alone Cannot...

Nov 7 · OPEN Tech Talks: AI wort...

32:32

Courses to attend

Red Teaming LLM Applications (DeepLearning.AI): hands-on attack techniques against your own LLM applications. The fastest way to feel, firsthand, what this week's Deep Dive describes from the attacker's chair.
AI Red Teaming training (Microsoft Learn): free guidance and labs from Microsoft's AI Red Team. Useful for turning this week's Monday checklist into a standing program your security team owns.

Events

Global AI Show, Abu Dhabi, November 12-13
Dubai Future Week, Dubai, November 18-21
AI Future Forum, Dubai, December 1-2
GITEX Global, Expo City Dubai, December 8-11
NeurIPS 2026, Sydney, Atlanta and Paris, December 7-13

Tool / Product Spotlight

Perplexity Teammates: always-on work agents that take standing assignments, monitor changes, and collaborate, inheriting your existing projects and permissions. You can inspect or redirect their work mid-run. Worth testing if you already run recurring research or monitoring by hand and want an agent whose work you can audit while it runs.
DigitalOcean Cuttlefish: an iPhone app that streams your live agent sessions and lets you approve or deny each tool call, steer a run, or kill it from your phone. It supports Claude Code, Codex CLI, LangGraph, CrewAI, and more. Worth testing if agents already run in your shop and this week's Deep Dive left you asking who supervises them at 2 a.m.

The Investment in AI

OpenAI raised an additional $11 billion at a $1.25 trillion valuation (TechCrunch), and inference provider Fal raised $1.3 billion at a $16 billion valuation, double its mark from earlier this year (TechCrunch). Read both against Thursday's $2.3 trillion public-market wipeout. In the same week public capital fled the AI story, private capital wrote two of the year's larger checks, and the Fal round priced raw inference demand, not narrative. The signal: the investors closest to usage data still see demand as real. What got repriced this week is who funds the build-out and on what terms, not whether the workloads exist. Write your own AI plan on your usage numbers, and let neither market's mood write it for you.

That's it for this week: thanks for reading!

If one idea here helps you, forward this issue to a colleague who makes AI decisions. Hit reply and tell me what you tried. I read every answer.

Until next Saturday,

Kashif

The views expressed here are solely my own, based on my experience, practice, and observation, and do not necessarily represent those of my current or previous employers or their clients and customers.