AcceleratingBiz®

Growth CEO’s AI Briefing · Edition #135

Check the Locks

Tuesday, October 20, 2026 · 3 Developments · 2 Perspectives · 1 Case Study · 2 Resources

Designed by AcceleratingBiz®. Curated, synthesized, and produced by Claude agents. AI-assisted production means occasional errors may appear.

Today's reading is about locks. Criminals are pointing cheap agents at ordinary security gaps, Washington is policing which models sit inside the software it buys, and the companies running agents well are writing permission rules into code.

This week, ask for a list of every agent with access to your systems and what each one is allowed to do.

Developments

Technology

Criminal AI Agent Fleets Broke Into Online Stores for About $50 a Success, and Left Their Dashboards Open (October 19, 2026)

Security startup NorForge published findings on October 19 from more than 20 exposed control panels for Cairn, an open-source agent tool, that it assesses as criminal. One cluster of 14 panels ran over 1,100 attack projects against online stores and payment systems, mostly in the US, Japan and Australia; NorForge estimates about 10% succeeded, each taking just under 23 hours. At roughly $5 in model fees per attempt, that implies about $50 per break-in, and worker names point to Claude Code running DeepSeek V4.1 Flash. Press reports cited by The Neuron and The Rundown named targets including John Deere, SharkNinja and Cummins; SharkNinja says the threat is contained.

So what

The agents' to-do list was ordinary: login flaws, pages that skip authentication checks, and file uploads. Cheap automation means known holes in your storefront, customer portal or admin pages will be found faster, so ask your security lead this week for a scan of exactly those three categories and a patch timeline. If you hire red teams, ask how they secure their own AI tooling, since NorForge found legitimate red-team panels exposing client data. Investors should add agent-driven attack exposure to portfolio security reviews.

Hype check

The 10% success rate and $5 cost are NorForge's own rough estimates, which it says could be off by a clear margin, and NorForge sells sandboxing tools for agents.

Confidence: Established fact — NorForge verified the exposed panels and one live back door; success and cost figures are estimates.

Go deeper NorForge · The Neuron

Legal & Regulatory

DHS Orders Civilian Agencies to Remove DeepSeek, Including Third-Party Products Built on It (October 20, 2026)

The Department of Homeland Security published an order in the Federal Register on October 20 barring federal civilian agencies from buying products from DeepSeek and its affiliates. Agencies must also remove existing DeepSeek products, and the removal extends to third-party products and services developed using DeepSeek implementations. CyberScoop reports the underlying order dates to June 26 and that exceptions require a request 60 days in advance with a compelling justification. The notice lands days after DeepSeek's V4.1 Flash update, which we covered Monday as a low-cost coding option.

So what

If you sell to federal agencies, or to contractors who do, a DeepSeek model inside your product is now a disqualifier rather than a cost choice. Ask engineering for a written inventory of every model in your stack, including those reached through routers, open-weight hosts and coding agents, and keep it ready for procurement questionnaires. Expect state governments and regulated enterprises to copy the language. Teams that adopted DeepSeek on price should line up a fallback model now.

Confidence: Established fact — the Federal Register notice is public; how far the third-party clause reaches will depend on agency enforcement.

Go deeper Federal Register · CyberScoop

Competition

Mercury 3 and MiniMax M3.1 Flash Push Fast, Cheap Models Further Down the Price Curve (October 19, 2026)

Inception released Mercury 3 on October 19, a diffusion language model that drafts and refines many tokens at once instead of writing one at a time. It claims 1,068 tokens per second and parity with GPT-6 Luna and Claude Haiku 5.5, at a launch price of $0.10 per million input tokens and $0.35 per million output tokens, half its standard rate. The same day, MiniMax launched M3.1 Flash at $0.15 input and $0.60 output, claiming top agentic and coding scores among Chinese models, with weights promised later. The Deep View argues open Chinese models are eroding US labs' pricing power at the low end.

So what

The tier of models that handles routing, extraction, support replies and voice gets faster and cheaper every few weeks, so last quarter's AI cost assumptions are probably too high. Ask your team to rerun one latency-sensitive workflow, such as live chat or voice support, on a fast model and compare speed, cost and error rate on your own data. Weigh the DHS order above before choosing a Chinese model for anything government-facing. Keep frontier models for hard reasoning, where these do not compete.

Hype check

Speed and benchmark claims for both models are the vendors' own. The Neuron's spot test of Mercury 3 found it accurate on math and code but over a word limit and prone to an unrelated error.

Confidence: Watch this — launches and prices are documented; performance claims await independent testing.

Go deeper Inception · The Deep View

Perspectives

AI Strategy & Transformation

Ethan Mollick After 30 Months of AI: Leaders Should Work as "Conjurers" Who Delegate to Agents (October 20, 2026)

Wharton's Ethan Mollick marked the October 20 release of his second book, Co-Existence, with a chart-based look at progress since Co-Intelligence in early 2024. He counts 180 models now better than GPT-4 was at launch, 19 of which run on a smartphone, and notes the METR long-task benchmark stopped measuring at about 17 hours of human work. The cost of a correct answer on a graduate-level science test fell from 30 cents to four hundredths of a cent. He adds a third way of working, the "conjurer," who summons agents, assigns work and judges the output, and he built the post's charts that way.

So what

Mollick's framing gives leaders a practical test: which work is still done chat by chat that could be delegated to agents and reviewed? Pick one recurring analysis, such as a board-pack section or competitive update, and have an executive delegate it end to end this month. Invest in what he says still favors humans, namely deep knowledge, wide knowledge, taste and agency, because reviewing agent output is where they matter. Keep his warning in view: delegation means less sight of how the work got done.

Confidence: Informed perspective — a respected researcher's synthesis of public benchmarks, written to launch his book.

Go deeper One Useful Thing · One Useful Thing archive

Technology

Microsoft's Security Chief: Treat Your Own AI Agents as Insiders That Need Watching (October 19, 2026)

Microsoft Security EVP Hayete Gallot argued on October 19 that agents deployed inside a company are "agentic insiders" that should not be implicitly trusted. Her reasoning is that models are trained to reach results, which can lead to reward hacking and actions outside company norms with no compromise at all. She proposes giving each agent its own identity, container and permissions, scoring the risk of each action in real time, and using supervising guardian agents to review actions before they run. The post cites no data and points readers to Microsoft's Project Perception.

So what

Many growth companies give agents the same access as the employee who set them up, which is the opposite of what Gallot recommends. Ask IT to list every agent with access to customer data, finance or production systems and confirm each has its own identity and the narrowest permissions it needs. Require an action log you can review after an incident. Today's Every case study reached the same conclusion from the operator's side.

Hype check

The essay doubles as a pitch for Microsoft's security products and offers no incident data.

Confidence: Informed perspective — a vendor's framework, consistent with standard least-privilege practice.

Go deeper Microsoft · The Neuron

Case Studies

AI Strategy & Transformation

Every Scrapped Personal Agents for One Shared Company Agent, and Moved the Rules Into Code (October 20, 2026)

Every's head of platform, Willie Williams, wrote on October 20 that the company deleted 86,730 lines of code from Plus One, its personal-agent product, after it proved hard to run and many users stopped using it. The tell was that people got the most value from borrowing a colleague's agent, so Every rebuilt it as one shared Slack agent on Anthropic's Claude Managed Agents. Shared use broke three assumptions: one conversation partner, one person's logins and one person's permissions. Fixes included tying scheduled tasks to swappable login "slots," putting permission rules in code the model cannot talk its way around, and asking first only when an action is hard to undo.

So what

If you are rolling out agents, one shared team agent may get more use than giving everyone their own. Before you do, decide who owns each workflow when someone leaves, whose credentials scheduled reports run on, and which actions need approval. Enforce access rules in the system, not the prompt. Every also found strict defaults locked teammates out of shared skills, so review permissions a week after launch.

Confidence: Informed perspective — a candid first-party account from a company that sells the product it describes.

Go deeper Every · Every Agent launch

Resources

Go-to-Market

Basis Survey: 48% of Marketing Teams Replaced Tasks With AI, but 65% Cannot Say Whether It Pays (October 19, 2026)

Ad-tech firm Basis surveyed 184 agency and brand marketers for its fourth annual AI and the Future of Marketing report, summarized by State of Brand on October 19. The share whose organizations replaced human tasks with AI in the past year rose to 47.8% from 27.9%, and 51.1% now run AI agents. Yet 65.1% either do not measure AI return on investment or do not know whether anyone does, and only 6.5% evaluate AI spend against defined benchmarks. Just 13% describe their first-party data as extensive and well structured.

So what

Marketing is cutting work on faith, which leaves the CMO exposed when the CFO asks what the AI budget bought. Ask your marketing lead to name one AI-replaced task and show its before-and-after cost, cycle time and output quality. Fix the data foundation in parallel, since agents cannot outperform messy first-party data. This adds the marketing view to Monday's Wharton finding that teams measuring growth outcomes report better returns.

Confidence: Informed perspective — a small, vendor-run sample of 184; directionally useful, not representative.

Go deeper State of Brand · TLDR Marketing

Customer Behavior

Kevin Indig Finds No Hidden AI Traffic in "Direct," Mostly Bots Instead (October 19, 2026)

Growth Memo's Kevin Indig examined Google Analytics 4 data for 12 mostly B2B sites and published the results on October 19. Direct sessions rose 141% from January to September, but engaged Direct sessions stayed roughly flat; one site logged 409,000 Direct sessions, of which 198 were engaged. Sessions with identifiable AI traces were about 0.003% of Direct, and 632 correlation tests between AI visibility and later traffic did not survive mismatched-timing checks. He concludes analytics alone cannot detect AI-driven demand.

So what

If your team credits rising Direct traffic to ChatGPT or AI search, check for bot spikes before it reaches a board slide. Add a required "how did you hear about us, and which AI tool?" field to demo and signup forms, and tie answers to pipeline and revenue. Track brand mentions across a fixed set of buyer prompts as a separate visibility metric. After last week's ChatGPT tagging bug, this is another reminder that AI attribution still runs on self-report.

Confidence: Informed perspective — careful analysis on a small sample of 12 sites with limited weekly data.

Go deeper Growth Memo · TLDR Marketing

Sources in this edition: The Neuron / Mindstream, One Useful Thing, The Rundown AI, The Deep View, Superhuman AI, Every (AI & I / How I AI), TLDR AI, TLDR Marketing, AINews (swyx)

← Latest Editions and Search