The Board Room

A new economic framework from MIT's Christian Catalini — surfaced through a16z — identifies what may be the single most important dynamic of the AI era: automation costs are plummeting, but verification costs are not. This gap is widening with every model upgrade, and the companies that fail to invest in verification infrastructure are quietly building toward catastrophic, visible failures.

The winners of the AI era won't be those who automate fastest. They'll be those who verify best.

The Data Is Now Concrete

Anthropic's Claude Opus 4.6 found 22 confirmed vulnerabilities in Firefox in two weeks — 14 high-severity — accounting for roughly 20% of Mozilla's high-severity bugs remediated this year. The cost: approximately $400 per vulnerability found. The cost to exploit: $4,000. That 10x finding-vs-exploiting asymmetry is today's ratio. Anthropic's own researchers explicitly warn it will shrink. OpenAI has responded by launching Codex Security as a free AppSec agent for enterprise users.

But defensive scanning is only half the picture. Multiple credible sources warn that AI agents are now pushing code into production repositories with less human review than ever. Factory AI's pipeline runs 40+ CI checks in under 6 minutes — the optimistic case. The pessimistic case: prompt injection spreading through high-profile open-source projects via agent-committed code that no human actually reads.

The Organizational Model Is Inverting

Catalini describes what he calls the 'AI sandwich': a single director, a swarm of AI agents, and a small team of elite verifiers. The traditional talent pyramid — many juniors, some seniors, few directors — is collapsing. Software engineering now accounts for over 50% of Claude model usage, and Citadel's labor data shows engineering postings rebounding even as overall white-collar roles decline. This is Jevons Paradox at work: AI makes engineers more productive, so organizations need more engineers to deploy agents across every other function.

The trap is what Catalini calls the 'codifier's curse' — your best people, by creating training data and evaluations, are automating away their own peers. This creates a dual crisis: a retention problem among your most valuable verifiers today, and a pipeline collapse in 3-5 years when no one has apprenticed into senior judgment roles.

The New Moat: Proprietary Failure Data

If your competitive defensibility relies on network effects from human coordination costs — the classic marketplace play — AI agents can cheaply bootstrap both sides and erode it. The new defensible asset is proprietary failure data: a decade of production errors, edge cases, near-misses, and failure modes that enable superior verification and risk underwriting. If you have this data, you're sitting on an asset potentially worth more than your product. If you don't, your moat is thinner than you think.

Catalini also introduces 'liability as software' as an emerging category: as AI agents function as autonomous workers, insurance and liability quantification become integral to software production. This market barely exists today but will be table stakes within three years.

February's jobs report didn't just miss expectations — it confirms a structural labor market contraction that has been hiding in revisions for three months. The -92K headline is bad. The December revision from +48K to -17K is worse, because it means every forecast built on Q4 2025 data was wrong. Unemployment ticked to 4.4%, with specific contractions in IT, manufacturing, and government payrolls.

The Fed is trapped: rising energy costs make cuts inflationary, but rising unemployment makes holding rates destructive. Officials explicitly signaled they need 'more alarm bells' before acting.

The Energy Shock Compounds Everything

Crude at $90.77 (up 12%) and gas jumping 30+ cents in a single week are input cost shocks that cascade through every line of a P&L. Trump's demands for Iran's 'unconditional surrender' while Israel enters a 'new phase' of strikes offers no visible off-ramp. Defense sector ramping to 'quadruple production' of weapons systems signals that federal fiscal priorities are shifting toward military spending — potentially at the expense of civilian technology procurement.

The Physical Infrastructure Bottleneck

The AI buildout's most underappreciated constraint is physical labor. The data center construction boom needs 300,000+ electricians at wage premiums of 4.3x the national average ($130/hr vs. $30/hr). Companies are building $700B in luxury temporary housing — complete with steak dinners and golf simulators — to attract tradespeople to remote sites. This means compute will remain scarce and expensive for longer than consensus expects.

What This Means for Your Plan

The macro picture is compounding headwinds: weakening labor market, rising energy costs, paralyzed Fed, geopolitical escalation, and a physical infrastructure bottleneck for the one sector supposed to drive growth. This is not 2022's rate shock — it's a multi-vector stress test. IT-specific payroll contraction adds fuel to the AI-displacement narrative, though pandemic overhiring is a competing explanation. Either way, the practical implication is identical: your enterprise customers are entering budget compression mode.

Companies that secure compute capacity now — through long-term contracts or direct infrastructure investment — will have a durable competitive advantage. Companies waiting for the market to normalize will wait a long time. The $166B tariff refund saga (Supreme Court struck them down, CBP needs 45 days, government is appealing) creates additional corporate balance sheet uncertainty through at least mid-Q2.

Two events this week, taken together, reveal that the infrastructure you trust most — your network perimeter and your MFA — is compromised at an industrial scale that demands immediate action, not next quarter's review cycle.

Cisco: Not a Bug, a Pattern

Cisco disclosed 50+ vulnerabilities simultaneously across Catalyst SD-WAN Manager, Secure Firewall ASA, FMC, and FTD. This includes critical authentication bypasses at CVSS 10.0 and two CVEs under active exploitation (CVE-2026-20122 and CVE-2026-20128). This isn't a single flaw — it's a pattern across an entire product family that suggests architectural debt in critical infrastructure products. Patches were available in late February, but any organization still unpatched is operating with known, actively exploited vulnerabilities in core networking.

The strategic question isn't 'did we patch?' — it's 'should we be this dependent on a vendor with this vulnerability profile?'

Tycoon2FA: MFA Defeated at Industrial Scale

Europol seized 330 domains and disrupted Tycoon2FA — which, by Microsoft's own data, was the platform behind more than 60% of all phishing Microsoft blocked in 2025. A single phishing-as-a-service operation was the dominant source of credential theft against the world's most deployed productivity suite. The technique — intercepting session cookies and 2FA codes via proxy servers — means traditional MFA was being systematically defeated. The takedown is a tactical win, but the technique is proven, documented, and will be replicated.

The Attribution Signal Changes the Threat Model

GTIG's data on the 42 attributed zero-day exploitations reveals a critical reframe: 39% came from commercial surveillance vendors and 28% from state-sponsored actors — roughly 70% from government entities or their contractors. Cybercriminals accounted for only 21%. The most dangerous adversaries aren't ransomware gangs — they're well-funded, patient, and operating with state backing. Your defensive investment must match the caliber of your most likely adversary, not the least sophisticated one.

The AI Dimension

Bing AI promoted malicious GitHub repositories as its top search result for a popular download — AI search systems are now a confirmed malware distribution channel. As enterprises deploy AI copilots and autonomous agents that pull code and dependencies from the internet, the trust model underlying AI-recommended actions is untested against adversarial manipulation. This gap needs controls before your AI agent strategy scales, not after.