Sage Wohns

The through-line

Wohns keeps returning to a single claim: generative AI did not invent social engineering, it removed the economics that used to limit it. His background is in machine reading and writing at scale, and he dates the turn precisely to the moment GPT-3 arrived and one of his largest customers in the DoD "tried to classify it as a weapon because now you can socially engineer at a scale that's absolutely unprecedented with precision that's never been before humanly possible" [1]. He makes the same point about technological waves weaponising quickly and about human-centric attacks reaching an unprecedented scale in his more recent appearances [2][5][6]. The consequence he draws is about price rather than capability: "No one's worried about burning a $10 million zero day anymore when you can just for pennies create a Marco Rubio recording and distribute that" [1].

From that premise everything else follows. If attacks are cheap, personalised and infinitely iterable, the defensive stack built for templated phishing is obsolete, the population that needs defending is far larger than the enterprise market anyone has been selling to, and the most valuable asset in the field becomes data about what a high-efficacy generated attack actually looks like. His earlier work in applied AI for finance and news organisations [3][4], and his experience at Rakuten and Agolo in machine learning and natural language processing, are the foundation he now applies to cybersecurity [7].

On why the old awareness training no longer works

"The old approach of using templates, the old approach of doing like persona based attacks, it's not relevant anymore" [1]. Wohns argues that attacks are no longer keyed to org-chart position alone. They are built from breach data, leaked PII and details such as which bank a target uses, which pushes everything from phishing to spear-phishing at the individual level [1]. The training that accompanied the old tooling fails for the same reason: off-the-shelf videos are "necessarily generic because they didn't have the ability to create dynamic targeted education" [1].

He is equally dismissive of the detection heuristics people have been taught in succession. The instruction used to be to spot grammar and spelling errors, then it became to spot a deepfake, and neither survives a world where most communication is AI-enhanced and everyone drafts emails, LinkedIn posts and resumes with the same tools [1]. His formulation is blunt: "just spotting something that generative does not equal an attack. What really equals an attack is an attack" [1]. What he wants people trained on is the attack path and kill chain, which is patterned and therefore learnable, and which can be tied back to the specific data the attacker used to target them [1].

On red teaming that feeds education

Jericho's structure inverts the usual sequence. "The attack is the beginning point, but the education is really where we can add a layer of data as well as enrichment" [1]. Wohns describes a red team that trains a blue team: the blue team analyses attack information to tell a user what to look for, and the education arrives with context, showing "here's your breaches, here's where you're vulnerable, and this is how to protect yourself from these" [1]. That education can be tailored and produced in under an hour for an individual, explaining why they were targeted this way and how to stop it, and it extends to protecting colleagues as well as oneself [1]. The company's use of AI to identify and counter novel generative phishing attacks is the technical core of this [7].

The channel coverage has widened fast. A year before the conversation the product was email only, which he treats as the incumbent's blind spot, and it now spans cross-channel messaging, voice and video [1].

On the attack dataset as the real asset

Wohns describes holding "the largest data set of generated attacks in the world, and it's annotated by humans," calling it "the most valuable data set I've ever seen" [1]. Its first use is reinforcement learning for higher-efficacy attacks, which keeps the red team ahead. Its more interesting use is defensive: a triage unit trained on that attack data, and the ability to train novel defences that stop attacks in their path rather than only preparing users to recognise them [1]. He also sees the dataset opening onto attack origination and automated takedowns, all of it grounded in understanding what a high-efficacy attack looks like in this new age [1].

On an 85 to 90 percent unvended market

His macro read is that this category is barely served. "This market is still 85 to 90% unvented... it's absolutely stunning how underserved this market is globally" [1]. He points to over a billion dollars of breaches across Singapore, Korea and Japan in a matter of months as evidence of both the exposure and the opportunity, and cites a crypto fund in Taiwan and a major enterprise deployment in Japan among users [1]. The old seven billion dollar market has grown substantially with the attack surface and the proliferation of attacks [1].

The answer he built for that gap is a freemium model, which he was surprised nobody else offered given the problem affects everyone [1]. The reasoning is that individuals and SMBs were priced out of solutions that were previously gated, while direct-to-consumer cyber insurance is growing and often carries a requirement to use one, so the constraint is self-service rather than demand: "one click to set up a a full year of uh campaigns in less than 20 minutes" [1]. He is direct about the accessibility problem in security, where complexity and vendor sprawl deter the people who most need coverage [1].

On protecting people outside the enterprise

The question he says he gets most at speaking engagements is about grandparents and parents [1]. Where an aging population once faced a stack of postal mail, it now faces emails, SMS and voice calls, and he singles out Japan and Korea as places where demographics and attack volume intersect badly [1]. He treats this as a serious segment rather than a sentimental aside, precisely because these attacks are increasingly nation state sponsored and the victims have no professional reason to know what to look for [1].

On dual use and national security

Wohns frames the company's mission as protecting the world from threats created by generative AI, and describes founding it as "a moral imperative" once he had confirmed the size of the opportunity [1]. He notes Jericho is the only dual use technology in the field, and argues that alignment with national security is imperative to winning what he calls this war [1]. The threat data he cites supports the framing: attack proliferation over 4,000 percent, voice phishing up over 400 percent in the first half of 2024, and the North Korean campaigns targeting hiring that dominated conversation at Black Hat, all powered by the same core technology [1].

On building companies with AI

The through-line from his earlier work is that AI compresses timelines. He has discussed how AI is transforming product creation, collapsing processes that once took years and enabling faster, more adaptive business models [7], and he has spoken about combining programming with business strategy to move AI into mainstream use in finance and news organisations, and about what the founder role demands and the role models he looks to [3][4][7]. His path runs from building websites for local businesses as a teenager in Seattle through Rakuten and Agolo, where the machine learning and natural language processing foundations he now applies to cybersecurity were formed [7]. He observes that the same acceleration cuts both ways in the market, since products can now be built very quickly by very small teams [1].

Takeaways

• The decisive change is cost, not capability: generative tools mean attackers no longer need to burn expensive zero days when a convincing synthetic recording costs pennies [1].
• Template and persona-based phishing simulations are obsolete, because attacks are now assembled from breach data and leaked PII at spear-phishing granularity [1].
• Stop teaching people to detect AI-generated content: "just spotting something that generative does not equal an attack. What really equals an attack is an attack" [1].
• Jericho runs a red team that feeds a blue team, so simulated attacks generate contextual education tailored to a specific person's exposure and deliverable in under an hour [1].
• The human-annotated dataset of generated attacks powers both higher-efficacy red teaming and novel automated defences, with attack origination and takedowns as the next step [1].
• Wohns puts the market at 85 to 90 percent unvended and answers it with freemium self-service, targeting SMBs, cyber-insurance-driven demand and consumers, including aging populations in Japan and Korea [1].