AI Barbed Wire

Schumpeter's coinage for the emerging category of infrastructure firms selling protection, trust, control and containment for AI agents — as distinct from picks-and-shovels chip/compute providers (Nvidia et al.). The name analogises to Alan Greenspan's argument that barbed wire, not railroads, was the technology that let settlers follow the frontier — because it made property defensible.

Why the category exists

The Autonomous AI Cyberattack class (4 labs in ~2 weeks per Should AI Labs Be Treated Like the Owners of Dangerous Animals (Economist)) + the everyday-blunder side (invisible errors that plausibly-pass, Agentic Exfiltration via Lethal Trifecta) together produce an enterprise-adoption stall on top of the cost stall the vault tracks via the cost-stall page. Jared Sine (GoDaddy) named it cleanly for the vault: "All you have to do is get snake-bitten once and you'd never go back."

The five sub-categories

Per AI Agents Lie Cheat and Steal (Economist):

AI-capable cyber-security — Palo Alto Networks · CrowdStrike (share prices ~doubled YTD); Alphabet-Wiz $32bn; >$70bn cyber-M&A in 12 months.
Agent trust-layer / data-leak prevention — Cyera ($12bn valuation, quadrupled in 18 months): prevents data leaks + unauthorised tool use.
Guaranteed-reliability model training — Scaled Cognition (co-founded by Dan Klein, Berkeley; $100m VC): reliability trained into the model, targets invisible errors — plausible-enough-to-miss errors that compound across long agent tasks.
Agent-identity + liability providers — startups assigning identities to agents so liability lies with a specific agent when it goes rogue. Market-based complement to Strict AI Liability on the legal side.
Kill-switch + control systems — vendors of the technical primitive tracked separately as Kill Switch.

Where this sits in the vault's AI-governance map

The vault previously carried three axes of response to the Autonomous AI Cyberattack class:

Axis	Instrument	Vault anchor
Eval-side self-regulation	Lab-administered pre-release evals	Hassabis FINRA-model proposal
Legal-side post-hoc liability	Strict-liability regardless of intent (wild-animals doctrine)	Strict AI Liability
Alignment-side R&D	Grant-funded alignment research	Coefficient → Resolution $160m July 2026

AI Barbed Wire is the fourth axis: commercial-market-supplied enterprise-side defence — priced by cyber-firm share moves, agent-trust-layer valuations, and reliability-model VC rounds. Distinct from the three above because it lives inside the enterprise's own deployment boundary, not at the lab or the courtroom or the grant-funded research programme.

The "harness = barbed wire" reframing

Schumpeter's closing line — "One of this year's AI buzzwords is 'harness' — the system that surrounds an LLM to keep agents on the straight and narrow. It might just as well be barbed wire." — reframes the vault's Harness (LLM Agents) page from a productivity frame (Boris Cherny / Ryan Lopopolo harness-engineering) to a containment frame. Same primitive, two different design goals — worth carrying because enterprise buyers will price both.

Follow-ups

First named enterprise-scale kill-switch vendor — the sub-category is real but the vault doesn't yet carry a named firm.
First agent-identity provider to publish liability terms — the market-based counterpart to Strict AI Liability becoming operational.
Cyber-security firm AI-agent-defence revenue split — Palo Alto + CrowdStrike H2 2026 earnings should surface the AI-agent-attack-mitigation line item (if they name it).
First insurance policy priced against the Strict AI Liability regime — Artificial Intelligence Underwriting Company is the leading candidate.
A cost-stall-side counterpart — whether barbed-wire spend adds enough to the enterprise AI-adoption bill to become its own cost-stall driver.

Cross-references

AI Agents Lie Cheat and Steal (Economist) — the anchor Schumpeter piece
Autonomous AI Cyberattack — the phenomenon this responds to (rogue side)
Agentic Exfiltration — the phenomenon this responds to (as-designed side)
Strict AI Liability — the legal-instrument sibling
Kill Switch — the technical primitive one sub-category vends
Harness (LLM Agents) — Schumpeter's harness-as-barbed-wire reframing
Companies Are Scrambling to Curtail Soaring AI Costs (Economist) — the cost-side stall this trust-side stall runs alongside
Cyera · Scaled Cognition · Wiz · Palo Alto Networks · CrowdStrike · Artificial Intelligence Underwriting Company — named firms in the category
Dan Klein · Jared Sine · Dawn Song — named voices