Traditional enterprise security is built for humans clicking buttons. It fails catastrophically when AI agents start managing workflows, according to Maxim Bar Kogan, CEO of Onyx Security. A system authorized to manage a database can see a ‘delete’ command as legitimate, even if the agent hallucinated the request. Identity management tools lack the context to interpret intent. Enterprises are granting agents broad access for productivity, creating a massive, autonomous error surface.
Kogan argues this is not a problem labs like OpenAI or Anthropic can solve. They are the car dealers; enterprises need an independent safety inspector. Firms are unwilling to hand historical interaction data back to labs that might use it to train future models. Onyx’s solution uses small, specialized ‘not smart’ models to flag anomalies in real-time, reserving expensive frontier model analysis only for critical moments. It mirrors blitz chess: intuition for 90% of moves, deep calculation for the high-stakes ones.
“If I pay $1 for a task, I can’t pay $2 for a second LLM to watch the first. That’s just a financial non-starter.”
- Maxim Bar Kogan, No Priors: Artificial Intelligence | Technology | Startups
The economic model for this new oversight layer depends on an agent economy that itself is breaking the internet. Cloudflare CEO Matthew Prince predicts AI agent traffic will exceed all human traffic by early 2027. Agents don’t click ads. They strip-mine the web for data and never send a visitor - or a payment - back to the source. Publishers collapse, and the data AI needs to learn stops being produced.
Prince says the only fix is a payment layer capable of 100 million transactions per second, a scale no current blockchain supports. Until that plumbing exists, the internet remains trapped in an “advertising slum” agents cannot navigate. The call for an independent, high-speed payment system mirrors the call for an independent, high-speed security layer: both are foundational infrastructure the original internet failed to build.
“AI agent traffic will exceed all human internet traffic in the first half of 2027. The economic foundation of the independent web is collapsing.”
- Matthew Prince, Bankless
Institutional players are already building their own moats. Kirkland & Ellis, the world’s largest law firm, is committing $500 million to a private AI platform, with $100 million allocated this year. According to The AI Daily Brief, Chairman John Bailis sees general tools like Harvey as raising the industry floor. Kirkland is paid for the ceiling. The spend is a defensive maneuver against middleman risk - if legal AI vendors eventually serve corporate clients directly, firms without proprietary tech become redundant.
On the user side, privacy is becoming a currency. Milan, co-founder of NanoGPT, says Monero has been the platform’s most-used payment method for ten months straight. Users funding ‘agentic’ workflows - software that manages its own wallet to buy compute - are removing the human from transaction logs. They are also deploying local 500-million parameter models in their browsers to act as private memory layers, preventing large labs from building a searchable database of their lives.
The core challenge is a structural one. Labs compete on intelligence, enterprises demand privacy, and the entire ecosystem needs oversight that doesn’t kill productivity. The $500 million law firm build-out, the push for 100M TPS payments, and the rise of Monero-driven compute all point to the same conclusion: the old centralized models are breaking. The new ones are being built in parallel, by players who refuse to be captured.
