ÜBER DIESE EPISODE
⬥EPISODE NOTES⬥
For a full year, agentic AI was the pitch. This year the conversation shifted to whether it holds up once it is actually running in production, against real work. Vendors came armed with customer-sourced numbers rather than concept demos — hours returned per analyst per week, percentages of alerts dispositioned without human review, agreement rates measured against human analyst judgment. Buyers arrived having spent the months since the previous major conference testing products and weighing what they were told against what they saw.
Not one of the four questions that dominated the show is exciting, and not one of them is actually an AI question. Naming an owner before something ships is governance. Showing your work is audit. Knowing where your components came from is supply chain hygiene. Configuring a tool to reach the outcome it was bought for is the oldest plain, unrewarded work there is. AI did not create those problems — it made them impossible to keep deferring. Marketing volume held steady. Scrutiny went up.
In this edition of Lens Four:
🔹 Why the easy read on Black Hat USA 2026 — that AI arrived — is a year late, and what moving from pilot to production actually exposed underneath it
🔹 The 4 questions that replaced the whether debate: who is accountable, what is the evidence, where is the data coming from, and is the foundation configured to hold the weight
🔹 How production turned governance into a named, accountable owner assigned before an agent ships, with a documented business justification behind it
🔹 The sharpest reframe of the week: a rogue agent is usually not malfunctioning, it is doing exactly what it was told, relentlessly, until it succeeds
🔹 The Midnight in the War Room session on the gap between people authorized to take risks and people expected to mitigate them, and why burnout rather than obsolescence is the analyst risk to manage
🔹 Why "black box" became unacceptable, and the 2 ways teams are validating: reading the reasoning trail directly, and replaying historical alerts against what human analysts already concluded
🔹 Sovereignty getting repaired in real time, from a geography and compliance word into a control word about who owns the derivative value of an organization's data
🔹 AI's own supply chain, and the 2 capability gaps leaders keep naming: a basic AI inventory, and third-party visibility into which vendor products already have AI embedded in them
🔹 The Vulnerability Research in the Agentic Age keynote on a pipeline producing well over 1,000 potential local privilege escalation findings, and why discovery got cheap while absorption did not
🔹 Post-quantum timelines compressing from 10 to 15 years toward as little as 3, and why crypto-agility belongs in the refresh cycle rather than a standalone initiative
🔹 What buyers were actually shown: roughly 75 percent of 10,000+ daily alerts cleared without primary analyst review, up to 19 minutes returned per analyst per hour at 99.7 percent agreement with human verdicts, and analyst throughput moving from about 10 closed alerts per shift to 50 or 60, all vendor-reported rather than audited
🔹 The 47 AI SOC vendors counted on the floor, roadmaps compressing from 12 to 18 months down to 3 to 6, and why the market got louder exactly as buyers got more specific
🔹 Why token costs that are not falling set a ceiling on adoption pace that governance readiness cannot lift, and what that means for consumption-based pricing
🔹 Why a resurgence of value-added resellers and system integrators is the market pricing a job that has to happen somewhere: which products fit which environments, and how you connect it all
Fourth Lens: Proof does not transfer. Different analysis approaches expose different properties, which makes tool efficacy hard to compare in the abstract, and prioritization frameworks are already moving toward environment-specific attributes. A number on a vendor slide came out of someone else's alert mix and someone else's data — it is evidence of something, but it is not evidence about you. So the burden lands partly on the buyer, with a split most people get wrong. The vendor owes the apparatus: a visible chain of reasoning, audit logs, the ability to replay your own history, and hands-on access without a six-month procurement cycle in front of it. The buyer owes the environment and the baseline. Neither side can produce the answer alone, and human on the loop is what that bargain looks like once it is running. If scrutiny only works when you have built something capable of doing the checking, what have you built?
▶ Full article and references: seanmartin.com/lens-four
▶ All Black Hat USA 2026 conversations: ITSPmagazine podcasts playlist
▶ Subscribe to Lens Four: seanmartin.com/lens-four
▶ Redefining CyberSecurity Podcast: redefiningcybersecuritypodcast.com
▶ Music Evolves Podcast: musicevolvespodcast.com
▶ ITSPmagazine: itspmagazine.com
▶ Studio C60: studioc60.com
Sean Martin, CISSP is co-founder of ITSPmagazine and Studio C60, host of the Redefining CyberSecurity Podcast and the Music Evolves Podcast, and the author of Lens Four, a weekly column on business, innovation, and messaging at seanmartin.com.
Keywords: Black Hat USA 2026, agentic AI, AI SOC, security operations, non-human identity, agent accountability, named accountable owner, rogue agent behavior, AI governance, chain of reasoning, human on the loop, AI supply chain, AI inventory, sovereignty, post-quantum readiness, consumption-based pricing, proof of value, CISO decision making
