The Board Room

The Cohesity Case Study Changes the Math

When a CIO at a $2B+ revenue company publicly states he can halve automation add-on spending using AI agents — and backs it with specific examples — the signal-to-noise ratio is exceptionally high. Cohesity's CIO replicated ServiceNow's ITAM tool in under 48 hours using Claude Code and has already replaced Splunk's SIEM with a custom AI agent. His projected outcome: 50% cuts to automation add-on spend while retaining core platforms for 1-2 more years.

The precision of this attack vector is what matters. Enterprise buyers aren't ripping out core SaaS — they're eliminating the add-on layer that carries 85-95% gross margins and drives net revenue retention.

This isn't an isolated experiment. Multiple CIOs across industries are independently converging on the same playbook: retain core platforms, eliminate add-on spend, build AI agents for process automation. The build-side cost just collapsed by an order of magnitude.

The Budget Reallocation Is Now Quantifiable

The macro data confirms the micro case studies. IT budgets are growing 3.4% overall while AI spending surges 81%. Anthropic and OpenAI now capture $40-50 billion annually from enterprise customers. This is near-zero-sum: CIOs aren't adding AI on top of existing software budgets — they're carving it out of SaaS expansion revenue.

Ramp's spending data quantifies the vendor-level shift: Anthropic captured 73% of first-time enterprise AI spend, up from roughly 50/50 just ten weeks ago. A 23-point swing in ten weeks in enterprise procurement — which typically moves glacially — implies either a catalytic product moment or accelerating disillusionment with alternatives.

Incumbents Are Signaling Capitulation

Salesforce issuing $25 billion in bonds to fund a $50 billion share repurchase is not a growth story. When the best use of $50B isn't R&D, acquisitions, or new market entry but buying your own stock, management is telegraphing limited organic growth opportunities. Compare to SAP, which is simultaneously executing buybacks and building aggressive AI partnerships with NVIDIA and Foxconn — playing both financial engineering and product reinvention.

Meanwhile, Workday's $1.1B Sana acquisition sets a new floor for AI capability M&A premiums, and Accenture's Q2 beat — forecasting AI partner work more than doubling in 2026 — confirms enterprise buyers are moving from pilots to scaled procurement.

The defense ServiceNow articulated — compliance, integrations, auditability — is real but narrowing. It holds in heavily regulated industries and erodes everywhere else.

The Velocity Trap Has Data Now

The industry's most sophisticated engineering organizations are confirming what scattered reports have suggested for months: AI coding velocity is actively degrading software quality and security. At Anthropic — the company that built Claude — 80%+ of production code is AI-generated, and it's creating critical UX bugs impacting millions of users. Amazon, arguably the most operationally disciplined engineering organization on the planet, has mandated senior review of all AI-assisted code after a rise in severity incidents.

Engineers classified as AI coding 'power users' produce 52% more pull requests, but downstream effects — outages, technical debt, security breaches — erode the value of that throughput. One analysis found teams spend 25% of their week fixing AI-generated code. This is the classic velocity trap: optimizing for output while degrading outcomes.

Any leader citing AI productivity gains to their board without accounting for the hidden velocity tax is building strategy on inflated numbers.

The Security Dimension Is Worse

GitGuardian data shows a 34% year-over-year surge in leaked secrets driven by AI-assisted coding. Claude Code commits leak credentials at 3.2% — more than double the 1.5% human baseline. Nearly 29 million credentials are exposed on GitHub, AI service credential leaks jumped 81% YoY, and 64% of valid secrets from 2022 remain unremediated.

Compounding this, Snowflake Cortex AI demonstrated a prompt injection vulnerability where an attacker escaped the sandbox, executed malware, and exfiltrated data using the victim's own credentials. This class of vulnerability affects every major AI agent platform. A separate concept — 'comprehension debt' — captures the organizational risk: when teams ship 3x more code but understand 40% less of what's in production, they haven't increased velocity; they've deferred risk into systems their own engineers can't debug.

The Counter-Signal: Autonomous Security Agents Work

Against this crisis, a genuine solution is emerging. Ramp's multi-agent security pipeline autonomously found and fixed approximately 100 novel security issues in six days with zero human involvement — using a coordinator agent, adversarial manager (40% false positive reduction), validator writing integration tests, and a fixer generating patches. Cursor prevented hundreds of production security issues in two months. OpenAI's Codex Security goes further by starting with architectural understanding and proving exploitability through micro-fuzzers.

Three independent approaches, all converging on the same conclusion: the future of application security is autonomous. The traditional SAST market faces its Kodak moment. But the gap between organizations deploying autonomous defenses and those still running human-speed security review is widening with every sprint.

The Stryker Attack Is a New Category of Incident

Iranian-linked group Handala didn't exploit a zero-day. They compromised Microsoft Intune — a legitimate MDM platform — and used its built-in remote-wipe functionality to factory-reset 200,000+ devices across 79 countries, claiming 50TB of data exfiltration. This is living-off-the-land doctrine applied to the management plane, and it changes the math on what constitutes critical infrastructure inside your organization.

Your MDM platform, your SSO provider, your OAuth integration layer — these aren't support systems. They're Tier 0 assets with destructive capability that rivals any malware.

The same week, three CVSS 9.8 FortiGate vulnerabilities were actively exploited via SAML token forgery — another management plane attack granting admin access through cryptographic verification flaws. And the 2025 Salesloft Drift breach showed OAuth tokens from a single vendor cascading to 700+ organizations, including Cloudflare and Palo Alto Networks.

Cisco's 3-Year Blind Spot Confirms Systemic Failure

Cisco disclosed 9 vulnerabilities, 5 actively exploited, with 2 SD-WAN zero-days exploited for at least three years undetected. The Interlock ransomware group was exploiting a max-severity Cisco firewall management flaw weeks before public disclosure. Researchers characterized edge infrastructure as 'prime real estate' for adversaries — but understated the implication: if your management plane is compromised, your entire security architecture is built on a foundation the adversary controls.

Layer on this week's SANS data: 80+ CVEs at CVSS 9.0+ in a single week, spanning security tools themselves (Wazuh SIEM root RCE, Veeam Backup five critical RCEs, ConnectWise ScreenConnect auth bypass) and AI platforms (Microsoft Semantic Kernel CVSS 9.9, SGLang CVSS 9.8, OpenClaw Agent Platform CVSS 9.8). When monitoring, backup, remote access, and endpoint protection all have critical vulnerabilities simultaneously, defense-in-depth becomes defense-in-name-only.

The Validation: Segmentation Saved Stryker's Revenue

Amid the destruction, one signal stands out as genuinely positive: Stryker's medical devices — Mako surgical robotics, LIFEPAK, Vocera, SurgiCount — survived the 200K-device wipe because they were architecturally isolated from the compromised Microsoft environment. The corporate IT fleet was devastated. The revenue-generating medical devices were untouched. This is perhaps the most expensive real-world validation of network segmentation in recent memory.

A new legal dimension adds urgency: Marquis is suing SonicWall after a ransomware breach exposed 672K people's data, alleging the firewall allowed attackers to steal configuration backups. If this establishes precedent, vendors face product liability risk and buyers gain contractual leverage they've never had.