Consuloji: Deep Dives into Overseas AI日本語

The Economics and Hazards of Frontier AI: Half-Price Inference, Swarm Token Implosion, Sheet Integrations, and 35,578 Unauthenticated Apps (GhostShares)

Published: November 13, 2026

Listen or watch on YouTube

A technical analysis of how sub-$2 frontier inference economics, recursive parallel agent execution, direct spreadsheet collaboration, and the exposure of 35,578 vibe-coded apps reshape enterprise AI governance.

Disclaimer: This publication provides objective technical, architectural, and security analysis. It does not constitute commercial endorsement or investment advice.


1. The Paradox of Modern AI: Cheap Intelligence Meets Security Blindspots

The marginal cost of frontier-grade artificial intelligence is collapsing at unprecedented speed. With Google’s release of Gemini 4 Carbon priced at $2.00 per million input tokens and $10.00 per million output tokens, state-of-the-art reasoning has dropped below half the historical cost of frontier systems. Concurrently, native office integrations such as Anthropic’s Claude Sheets and Whiteboards enable autonomous models to directly manipulate business logic and live canvas assets in real time.

Yet this era of frictionless access has simultaneously exposed structural vulnerabilities in software architecture and budget governance.

A forensic audit conducted by security firm Claroc uncovered that 35,578 applications (94.7% of 37,571 online instances tested) built on xAI’s Grok Build Apps opened without any authentication on the public internet. Among the open applications were corporate sales materials, medical records, voter queries, and third-party API credentials accessible to any web browser without sign-in.

Simultaneously, engineering teams deploying autonomous multi-agent swarms (e.g., Google Antigravity Swarm) are confronting “token implosion”—a failure mode where recursive agent feedback loops consume tens of millions of tokens and thousands of dollars in compute within minutes.

When intelligence becomes cheap and software creation requires only natural language prompts, traditional architectural hygiene is easily abandoned. This analysis examines the technical drivers behind frontier cost reduction, the mechanics of runaway agent execution, the governance of spreadsheet integrations, and the systemic architectural flaws behind the GhostShares unauthenticated exposure.


2. Slashing the Frontier Inference Floor: Gemini 4 Carbon and TPU 8i Max Architecture

The New Baseline of Frontier Pricing

Historically, frontier-class models commanded between $5.00 to $10.00 per million input tokens and $15.00 to $30.00 per million output tokens. Gemini 4 Carbon resets this baseline.

Offering / Generation Input Price (/1M tokens) Output Price (/1M tokens) Primary Architectural Focus
Gemini 4 Carbon (Introductory) $2.00 $10.00 Active through November 30, 2026; cuts top-tier inference price in half
Gemini 4 Carbon (Standard Target) $4.00 $20.00 Permanent baseline, maintaining aggressive price-performance ratio
Legacy Frontier Baseline $5.00 - $10.00 $15.00 - $30.00 High reasoning capability constrained by steep recurring token overhead
flowchart LR
    A["Legacy Frontier Tier / Input $5 - $10 / Output $15 - $30"] -->|Vertical Infrastructure Integration| B["Gemini 4 Carbon / Input $2.00 / Output $10.00"]
    B --> C["Impact: Continuous Background Inference<br>& Massive Context Swarms"]

Infrastructure Optimization: TPU 8i Max, Optical Circuit Switching, and Dynamic Slicing

Achieving an 80% improvement in inference price-performance over previous generations required full vertical hardware integration. The underlying eighth-generation TPU systems (TPU 8i Max and TPU 8tx) introduce two key technological breakthroughs:

  1. Optical Circuit Switching (OCS):
    • Eliminates intermediate electrical packet switching between TPU cubes.
    • Micro-electro-mechanical systems (MEMS) mirrors dynamically direct optical signals directly between fiber lines, slashing thermal loss and latency during massive context retrieval.
  2. GKE Dynamic Slicing:
    • Decouples physical rack hardware provisioning from logical slice topology.
    • Workloads dynamically assemble TPU cubes into larger super-slices or carve them into isolated sub-slices on the fly, eliminating idle compute during agent execution spikes.

3. Swarm Coordination Traps: Token Implosion and Runaway Recursive Execution

Lower unit costs incentivize multi-agent orchestration architectures, such as Antigravity Swarm. Tasks are decomposed across specialized autonomous workers—such as researchers, implementers, and reviewers—coordinating via shared message boards.

flowchart TD
    Coordinator["Coordinator Agent"] --> Board["Shared Message Board"]
    Board <--> Worker1["Research Worker"]
    Board <--> Worker2["Fast Worker (Implementation)"]
    Board <--> Worker3["Code Reviewer (Gate 2)"]
    Board <--> Worker4["Security Challenger"]

The Mechanism of Token Implosion

While parallelization accelerates complex workflows, unconstrained swarms exhibit exponential token consumption:

  1. Massive Shared Context: If eight workers each maintain a 100,000-token workspace state, a single coordinated step ingests 800,000 tokens.
  2. Adversarial Infinite Loops: When an automated reviewer rejects a worker’s patch and demands iterative micro-refinements, the cycle can execute twenty or thirty times without reaching consensus.
  3. Budget Evaporation: Within ten to fifteen minutes, an unmonitored swarm can consume 20,000,000 tokens, triggering immediate API quota throttling and thousands of dollars in billing.

Engineering Safeguards for Autonomous Swarms

Production deployments must enforce three deterministic constraints:

  • Concurrency Limits (concurrency_limit): Hard-capping simultaneous active sub-agents to 3–5 processes.
  • Session Budget Caps (budget_cap): An immutable ceiling on monetary expenditure or aggregate tokens per task session, triggering an immediate kill switch upon breach.
  • Step-Count Circuit Breakers: Mandatory process halting and human escalation after a pre-determined cycle threshold (e.g., 15 turns).

4. Bidirectional Workspace Editing: Formula Protection in Claude Sheets & Whiteboards

The migration of autonomous models from external IDEs into direct office productivity suites—exemplified by Anthropic’s Claude Sheets, Whiteboards, and Interactive Connectors—introduces severe operational data integrity risks.

The Problem of Destructive Overwriting

In traditional workflows, humans copy table fragments into chat prompts and manually paste structured responses back. Direct API synchronization removes this friction by allowing models to read row schemas, calculate aggregations, and update cells directly.

However, enterprise financial sheets contain mission-critical business logic: nested SUMIFS, dynamic XLOOKUP references, and auditable historical transactions. When models are granted unconstrained write access:

  • Formula Flattening: Complex dynamic formula cells are frequently overwritten with static calculated values, silently breaking downstream calculations.
  • Historical Data Erasure: Previous fiscal periods are inadvertently replaced with newly generated forecasts.
  • Reference Disruption: Arbitrary row or column insertions invalidate cross-sheet range pointers, triggering cascading #REF! corruption.
flowchart TD
    subgraph Danger["Unsafe Direct Write"]
        M1["Autonomous Agent"] -->|Direct Overwrite| S1["Unprotected Spreadsheet / (Formulas flattened to raw numbers)"]
    end

    subgraph Safe["Hardened Protocol"]
        M2["Autonomous Agent"] -->|Write to Staging Range| S2["Protected Sheet / (Locked Formulas & Historical Cells)"]
        M2 -->|Generates Visual Diff| Review["Human Review & Approval Gate"]
        Review -->|Explicit Sign-off| S2
    end

Mandatory Governance Protocols

  1. Range-Level Write Restrictions: Constraining agent write scopes strictly to designated input cells, while applying immutable sheet protection to formula and historical ranges.
  2. Review & Approve Checkpoints: Enforcing a mandatory diff inspection modal where changes must be explicitly approved by human operators before committing to production documents.

5. The GhostShares Audit: How 35,578 Grok-Built Apps Opened Without Authentication via Missing RLS

The most acute security failure emerging from generative application development is the GhostShares vulnerability disclosed by Claroc.

sequenceDiagram
    autonumber
    actor NonTech as Business User
    participant Platform as Grok Build Apps Engine
    participant DB as Backend Database (RLS Missing)
    actor Public as Public Web / Scanner

    NonTech->>Platform: "Create a shared customer tracker for my 4-person sales team"
    Platform->>DB: Generates schema without Row Level Security
    Platform->>NonTech: Publishes live URL (Public by default)
    Note over NonTech: Assumes tool is internal because login UI appears
    Public->>DB: Sends unauthenticated direct REST query
    DB-->>Public: Returns entire table: customer PII, corporate logs, API keys

Forensic Breakdown of the Exploit

  1. Public by Default Architecture: The platform configured newly published applications as globally accessible on the internet by default, requiring explicit manual configuration to restrict access.
  2. Absence of Row Level Security (RLS):
    • When users prompted for a “team customer list,” the engine provisioned a relational database and generated a login form.
    • However, the underlying database engine lacked Row Level Security policies. The backend API endpoints accepted unauthenticated HTTP requests, returning entire datasets regardless of the caller’s identity.
  3. Data Scope: Claroc confirmed that 35,578 online application instances (94.7% of 37,571 tested) opened without sign-in. Open apps discovered during the audit contained sensitive materials including sales territory maps, consumer ad proposals, medical records, and private API keys.
  4. The Vibe Coding Trap: Because non-technical creators interact purely through conversational syntax, the underlying infrastructure, schema architecture, and network policies remain invisible, creating an illusion of privacy.

6. Production Hardening Playbook: Private by Default, Hard Budget Caps, and Review Gates

To safely harness the power of sub-$2 frontier intelligence, organizations must institute the following operational controls:

Enterprise Governance Checklist

Domain Control Mechanism Implementation Standard
No-Code & Vibe Apps Private by Default Enforcement Mandate organization-wide SSO; enforce automated scans for unauthenticated backend endpoints before public domain routing.
Database Security Mandatory RLS Policies Require table-level Row Level Security policies on every generative database, blocking anonymous SELECT/UPDATE queries.
Office Integrations Formula Locking & Diff Review Protect calculation ranges in Claude Sheets; mandate human approval gates before changes are written to disk.
Autonomous Swarms Hard Budget Caps & Step Timeouts Enforce immutable token and cost budgets per task; implement circuit breakers at 15 recursive turns.

7. Conclusion: Governance as the Prerequisite for Scaled Autonomy

The radical deflation of inference costs and the emergence of direct office collaboration represent a fundamental leap in software capabilities. Intelligence is no longer a scarce, metered resource; it is an ambient, continuous execution layer.

However, as demonstrated by the GhostShares audit and multi-agent token implosions, automated creation does not eliminate foundational software engineering rules. Without row-level security, budget circuit breakers, and cell protection, rapid deployment merely scales systemic exposure.

Sustainable AI adoption requires building rigid security and cost governance guardrails before unleashing autonomous agents into production environments.

Back to articles