CVEs Are Going TO THE MOON
In a Nutshell
AI agents are already delivering 5x productivity gains at companies like Anthropic and Customers Bank, where code commits per engineer surged while headcount dropped 15% at Meta with 46% productivity increases. The explosion of CVEs represents a critical 6-12 month window where AI will identify and patch decades of vulnerabilities in human-written code, forcing organizations to refactor software infrastructure. The decisive factor is access to frontier models versus basic tools, with only 13% of companies actively redesigning work for AI—those that don't will fall permanently behind as agents become capable of operating as average digital employees within three months.
These notes were generated by AI and may contain inaccuracies.
Jensen Huang of Nvidia stated that AI will never replace domain expertise, though this claim has generated polarized reactions online. Some view this as the desired narrative of AI as augmentation rather than replacement, while others see it as disingenuous messaging. The comfortable lie contrasts with Anthropic's messaging where a representative told grad students their career plans went up in smoke.
Meta announced Q3 results showing 26% year-on-year growth while cutting 15% of staff. The cuts occurred across sales, product development, and moderation teams. Productivity increased 46% despite these reductions. Meta flattened the organization by cutting managers, and this approach appears sustainable given their frontier position in organizational redesign.
A traditional rule of thumb from Cisco systems indicated tech companies need 7% annual churn rate. Meta's 15% cuts exceed this average but align with Mark Zuckerberg's history of major pivots, including the expensive Facebook-to-Meta transition. The purpose of automation is labor-saving technology enabling growth with same or fewer headcount. Software represents AI's current strength, making it the first domain to saturate with automation.
Anthropic's economics team published a study on October 16th examining their own teams' effectiveness. The study analyzed internal employment data, tool usage, outputs, Slack, calendars, and hiring data across 17 interviewed employees.
Key findings:
- Code commits per engineer increased five times compared to six months earlier
- 192 engineers adopted coding agents versus 74 who did not
- Productivity gains extended beyond engineering to lawyers processing contracts, accounts payable handling invoices, and sales closing deals
- Quality remained consistent between agent-written and human-written code
- 80% of agent interactive sessions are idling, making human review the bottleneck
The junior share of technical hires at Anthropic fell from 27% to 9%. Junior share of employment overall dropped from 21% to 14% since early 2025. The information sector in America declined from approximately 3.2 million jobs to 2.7 million jobs following the post-COVID ZIRP rebound.
Historically, when farms industrialized, young men moved to cities for factory work. Current tech job drying up mirrors this pattern. Companies enabling junior employees with AI reach effectiveness parity faster. Call center studies show junior staff benefit more from Copilot support than senior staff who already possess tribal knowledge.
AI can capture knowledge from retiring employees through extended interviews that transcribe tacit knowledge into formal systems. Capital One uses public repositories for shared context where employees contribute via agents. This approach requires careful handling as people react negatively to programs explicitly framed as self-replacement.
Only 1% of companies leading in AI and automation are actively cutting headcount. Companies achieving the highest EBITDA growth focus on making everyone more capable rather than reducing staff. The standard deployment question focuses on cost savings and automation, but the growth question asks what could be accomplished with unlimited resources and super intelligence.
Critical Path emphasizes senior executives need exposure to frontier models to generate creative growth ideas. Most organizations deploy tools like Microsoft Copilot focused on optimization and productivity improvements. A minority introduce frontier engagements centered on opportunity, creativity, and growth.
October has been dubbed Black October in cybersecurity. CVE publication charts show parabolic growth, indicating an AI cybersecurity apocalypse.
When examining AI and cybersecurity, the surface-level analysis focuses on the increasing number of CVEs being published. However, what's not happening is actually more significant. Every single one of these vulnerabilities is being published by the good guys - Anthropic, OpenAI, Microsoft, IBM, and other American or American-allied organizations. The countries not discovering these vulnerabilities are North Korea, Russia, China, and Iran.
OpenAI and Anthropic frequently report that their agents went rogue and hacked websites. However, examination of these cases reveals they were deliberately unconstrained models where guardrails were turned off and explicit goals were given to compromise systems. The complaint that the AI broke the rules is circular since the AI was instructed to break rules and compromise systems at all costs. OpenAI has now learned to air-gap these training and test sessions from the internet.
Sam Altman stated he is not confident they can catch every bug, while Dario Amodei shares similar concerns. Both seek regulatory capture to pull up the ladder behind themselves. The old talking point has returned that unconstrained cybersecurity models are needed for red team versus blue team exercises and white hat hacking.
A month prior, the speaker stated they would only be impressed once AIs found new hardware exploits like Heartbleed and Spectre/Meltdown. AIs did find something similar using cache timing in CPUs. An agent running on one virtual machine can time the CPU cache to communicate with an agent on another virtual machine on the same physical host without networks. This is a known class of exploit that can be patched, requires a shared protocol, and only works on the same hardware, making the blast radius relatively small.
These are the months where IT infrastructure people, software engineers, and cyber professionals will earn their keep. Reuters reported a campaign from Norway where a Norwegian firm called Nor forged identified an operator's working file exposed on the open internet. A researcher tasked agents using KERN, an open-source platform for coordinating agents, and Microsoft's Playwright automation tool. The agents were instructed to find ways to run code without signing in, scanned over 1,200 websites, and entered more than a hundred of them.
We are living in a civilization running on 60 years of human-written code with vulnerabilities that AI agents can now mathematically identify and exploit. We are fortunate that responsible people hold the powerful models and are identifying vulnerabilities, contacting companies through back channels, and providing time to patch. This is a critical period that might last 6 months to a year, where the world will live through a complete software refactoring.
IT operations professionals understand the complexity and threats but face pressure from businesses demanding new capabilities. Funding needs to go into this space to ensure protection over the next 3 months to a year. The expectation is that software will be completely refactored, redesigned, and patched within a year.
Stories about rogue AI feed into fear and legitimize not using AI. During a strategy meeting being recorded for AI processing, one person out of 15 objected, stating they didn't want to get recorded because the model could go rogue. This person never spoke during the one-hour meeting, so recording would have had no consequence. These stories open doors to conspiracy theories about AI recording being leaked to competitors.
A viral tweet states that 3% of all compromises in computing history happened in the last week. The tweet argues that when lock picks are outlawed, only outlaws will have lock picks. The lock picks will be Chinese open weight models if nothing else. There are two futures: one where only bad guys have lock picks and security holes go undiscovered until exploited, and another where everybody has lock picks and aggressively red teams their own software. The corpus of human code is huge but finite, and we have a million friendly robots to fix defects.
The only way out is through. New paradigms are scary because humans have chimpanzee brains that fear new and different things. There is also ego threat regarding identity and skill sets. Technology has always been evolving with the lesson that adapt or die is the rule. This is the next paradigm shift.
Customers Bank Corp in Pennsylvania reported third-quarter results where they used AI, agents, and automation to rebuild their commercial deposits onboarding process. After a 10-week pilot, complex commercial accounts now open in minutes. Cycle time improved 90%, loan underwriting cycle time fell 95%, and complex loan closing shortened from 30-60 days to 7 days. They saved 76,000 cumulative hours and unlocked more than $50 million in annual capacity. Revenue per employee rose 13%, pre-tax profit per employee rose 34% year-over-year. They have a team of about 60 people working on AI transformation plus a dozen forward-deployed engineers from partners.
Team A versus Team B framework resonates well, with Team C being people with their heads in the sand who aren't playing the game. The better framework might be Team A and Team Triple A, where people performing at the same level suddenly have team members rocketing above through AI engagement. The Anthropic story showed five times the productivity of engineers as evidence of this.
Access to frontier models and pro plans is becoming increasingly important. OpenAI's Dot has context across all projects, allowing direct queries about upcoming meetings, relevant emails, and plans without providing context. Vanilla ChatGPT requires context filling. The productivity lift from being plugged into emails, apps, projects, and history is major. Employees with access to advanced and expensive models have significant advantages in meeting preparation and inbox management.
Only 13% of companies are redesigning work to take advantage of AI, representing a missed opportunity. If watching this content, you are one of the 13% paying attention while most competitors haven't realized significant gains yet. This is the Team A mentality.
The primary job of a CEO is risk control, especially for successful companies making $100-500 million annually. Every CEO is risk-averse and looks at what competitors, friends, and parallel industries are doing, only doing the median thing. AI is not just productivity software like Microsoft Office but as radical a change as going from pre-internet to internet-enabled business or pre-personal computer to personal computer-enabled business.
Grok bots should be watched as Elon Musk builds the Macro Heart mega data center. For the last couple of years, Musk has discussed digital Optimus as a digital employee. The power and compute coming online by end of year will power Grok Bot, which should show capabilities akin to a digital Optimus by year's end. This will be the big thing for 2027 - creation and adoption of truly capable digital employees.
Both Dot and Claude can identify work that should be done, which is a game changer. The next step is identifying when humans should approve versus when AI should proceed autonomously. Humans sleep, creating delays. The problem is AI asks too much whether to build things when action is preferred. Trust needs to be given to AI that sometimes makes mistakes, similar to human assistants.
We are coming to the point where we need to trust agents similarly to hiring colleagues. Like self-driving cars, we trust them because they make fewer mistakes. There should be a threshold that slowly increases rather than giving complete access to credit cards and investment accounts. AI is becoming not only a chief of staff but also a spokesperson and credit card with per-transaction limits. People have started connecting Grokbot to their XMoney accounts.
People have started connecting Grokbot to their XMoney accounts. Elon created XMoney not just for creator payouts, but because XMoney to XMoney accounts operate through their own clearinghouse. This system bypasses the Visa, MasterCard, and Swift networks, avoiding the need for approval from traditional banking systems. This mirrors Elon's original vision with PayPal, which was eventually absorbed into the conventional banking infrastructure.
Delta Airlines and other major companies shut down agent-based commerce this week. Some viewed this negatively, but the rapid reaction indicates these Fortune 100 companies are actively building internal agent infrastructure. Not all companies are responding this way—some are simply shutting down access without preparation.
A Twitter user argued that companies tried agentic checkouts in the 1990s, but this comparison is incorrect. Those were simple APIs running on unencrypted HTML, fundamentally different from today's autonomous agents.
Agentic commerce is actively being developed. The approach mirrors race car driving, where brakes, steering, and gas are all necessary tools. Using brakes allows for better steering around corners rather than representing a slowdown.
The biggest current discussion point centers on safely deploying frontier models beyond basic tools like Microsoft Copilot in Office 365. Organizations are concerned about protecting secrets and customer data while harnessing autonomous agents. The gap between optimization tools and true frontier intelligence is becoming the critical conversation topic.
In three months, agents will be capable of operating as average employees for digital work. Organizations need to begin thinking through these deployment questions now to avoid being left behind.
Using basic ChatGPT ($20/month), Claude, or free versions versus frontier models ($100-200/month) represents a night-and-day difference in capability. Without access to frontier models, users cannot build intuition for what AI systems are truly capable of achieving.
The recommended approach starts from the top down. Every executive should use high-tier frontier model plans and experiment with frontier capabilities, including new agents. This familiarity must come first before organizations can effectively implement AI strategies. Some CEOs are not technical, but someone in their inner circle needs to be a power user who can demonstrate capabilities.
Clients frequently request exact installation instructions and ordering for AI empowerment. However, mindset represents 70% of the equation. Without intuition for AI capabilities, organizations cannot ask the right questions or understand what to implement.
Four essential components are required to extract value from AI systems:
- Frontier reasoning capability as the foundational requirement
- Context including environment, history, and relevant background information
- Memory systems to maintain continuity
- Quality of interrogation - the ability to construct appropriate questions based on topic knowledge
Improving inquisitive skills through Socratic questioning is critical. The difference between surface-level questioning and deeply mindful, thoughtful guidance produces dramatically different outcomes. Frontier model intelligence combined with context, memory, and quality interrogation determines results.
A membership program launching in the coming week or two will have four tiers:
- Free tier: monthly newsletter
- Tier one (Frontier): weekly newsletter
- Two additional premium tiers with details to be announced
Growth over automation remains the key principle. The mindset of optimism points toward a sustainable abundance future. The path forward requires collective effort, learning from each other, and using AI for good. The approach emphasizes going through challenges rather than around them.
Keep David Shapiro in your library
Save the videos and channels worth coming back to, and find them again in one place.
