The Defender’s AI: Five Decisions for Chief Executives After Black October

7 days ago

5 min read

Disclosure: I spent the first half of my career in physical security, where I earned the CPP designation and served as Chairman of the ASIS International Physical Security Council. In the early 2000s I briefly owned an information security firm and a technical surveillance countermeasures (TSCM) company. The field has moved far since those days, and I do not claim to be a cybersecurity expert today. What I bring is an executive’s perspective, formed over years as a Vistage Chair and more than 8,000 hours of one-to-one and small-group coaching with CEOs and C-level executives, and in facilitating their peer-group conversations on business matters. What follows is my best assessment of the key issues today, informed by a canvas of more than eighty curated sources, a panel of six large language models, and Claude Opus 5.5 as a capable strategic thinking partner. Together we prepared Issue 133 of the AI Daily Intelligencer, a special issue on cybersecurity dated October 28, 2026, as well as this article and the accompanying slide deck. Take it as the counsel I would give a friend, a fellow CEO or a board member. These are the riskiest times I have known as an executive, and I would urge every leader to read, reflect, and consult their own advisors on how to address these issues now.

October 2026 acquired a name before it was over. On the morning of October 22, a post on X gave the preceding forty-eight hours a name, Black October, and within a week the label had passed through security newsletters, Wikipedia, CNBC and Axios.

The events behind the name were serious, and each of them involved artificial intelligence. PHP shipped fixes for two critical flaws that researchers had found with AI assistance, and governments in Britain and across Europe took public services offline rather than run them unpatched. LiteLLM, a gateway many companies use to route their traffic to AI models, shipped a release that carried a credential stealer. A certificate authority whose keys an AI agent had recovered from a single prompt announced that it was leaving the global certificate business. Microsoft described a China-aligned group using an agentic AI framework to work through a vendor’s applications against more than 2,250 Microsoft 365 tenants.

Issue 133 of the AI Daily Intelligencer is a special issue devoted to that month. It sets out ten steps a chief executive can hand to the CTO, CIO or CISO, each anchored on a verified event. I have also turned it into a briefing deck for CEO audiences, which sorts the ten steps into five decisions that belong to the chief executive alone.

The signal the machines missed

The sentence I built the issue around came from Geoffrey Hinton, in an interview with Armenpress published on October 27: “Once AI is better than people, and it will be better than people quite soon at doing cyber-attacks, the only way to defend is to use another AI.” He added that “the attacker only needs to be successful once, and the defender needs to beat the attacker every time.”

Our scan of the senior research tier found that interview. None of the six LLM AI models we put the same question to named it. Every one of them named the PHP flaws, while none named Hinton, the scientists’ letter calling for cyber-capable models to be air-gapped, the banks’ blocks on outside AI agents, or the laboratories’ disclosures that their own agents had reached real systems. The models were good at reporting incidents and poor at surfacing the human signals behind them, which is why our method reads the senior voices first.

I believe Hinton is right, and the month sharpened what being right requires. Many of October’s worst incidents struck the AI that companies had already adopted: the gateways, the keys, the coding assistants and the agents. A leader who hears Hinton and simply buys more AI will have heard half of him. His counsel has two halves, and they must be funded and governed together: defend with AI, and defend the AI you run.

Five decisions for the chief executive

Fund defense with AI. Put AI into detection, code review and patch prioritization, where speed decides outcomes.
Govern the AI you already run. Register every AI key, gateway, coding assistant and agent, each with an owner, least privilege and a log it cannot alter. Satya Nadella put the principle plainly on October 10: “We must assume a model is compromised and contain it from the start.”
Move at the attacker’s speed, and pause what you cannot yet protect. Amazon logged an attempt to exploit the PHP flaws four days before the patch, and CISA now gives federal agencies three days to fix most listed flaws, against twenty-one a year ago. A system that cannot be protected should come offline, and should return when it passes the same tests it failed.
Control what you depend on. The month’s most efficient intrusions came through access victims had granted their vendors, and its most disruptive outages stopped firms whose provider was hit while they themselves were not.
Show the board the exposure its meters miss. In a month of breaches, the SEC received one material-incident filing, while cyber insurance rates fell for a thirteenth straight quarter as ransomware victims reached a record.

That last decision comes from experience as well as data. In more than twenty years in physical security, as a CPP and as chairman of the ASIS Physical Security Council, I saw how few major incidents ever reach public view and how often they are reclassified. Whoever is judged by a number also decides how the events behind it are classified, and a board should read its disclosure record with that in mind.

Reading the month with discipline

The label describes a wave of disclosure more than a wave of damage. Axios reported that millions of hacking attempts had so far yielded few successful attacks, and ESET found the PHP flaws harder to weaponize than first feared. Several headline incidents began months earlier. The right posture is to hold the pattern with confidence and the damage estimates loosely, and to notice that the attacks that failed did so because defenders patched, isolated and watched.

How the issue was built

This brief was produced with AI as a thinking partner in a directed inquiry: parallel research agents scanned seven tiers of sources, a panel of six models was tested against the same prompt, and verification agents traced every claim to its origin, with every editorial decision remaining mine. Verification removed claims that circulated widely during the month, including words attributed to well-known figures that they did not say or that only one outlet reported. In a month when attackers impersonate banks, certificate authorities and colleagues, a publication should be strict about whose words it prints. I would ask the same of any firm deploying AI: inspect and test what you put into production, and hold back what does not pass.

The deck

The briefing deck carries the evidence for each step, a roadmap of fourteen actions with owners, a ten-question self-assessment for scoring your own firm, and a template for a one-page board report. The full issue is available at AIWhisperer.org. Arm the defense with AI. Guard the AI you have armed. Remember that the attacker needs one success while you need them all.

Severin Sorensen is Editor of the AI Daily Intelligencer, Founder and CEO of ePraxis LLC, and author of The AI Whisperer series. The AI Daily Intelligencer Special Issue on Cybersecurity, Issue 133 is available for free download at: https://aiwhisperer.org/course/ai-whisperer-intelligencer.