The AI stack forked this week — pick your side before McKinsey does

On Monday, IBM fell 13%. Salesforce, ServiceNow, and Snowflake each dropped 4%. Oracle lost 4.5%. The Nasdaq was down 1.2%. That gap is the story.

What triggered it: OpenAI told investors that Salesforce, Workday, Adobe, and Atlassian revenues are its addressable market, and simultaneously announced multi-year distribution partnerships with McKinsey, Accenture, BCG, and Capgemini for a new enterprise agent platform called Frontier. A day later, Anthropic launched Claude Cowork with vertical plugins for finance, engineering, and design, and integrations into DocuSign, LegalZoom, and Salesforce. Anthropic's head of economics framed the strategy in one sentence: replace the worker, not the software. The market read it immediately — Figma popped 10%, Salesforce reversed to +4%, ServiceNow ticked up.

Same week. Two opposite theories of how enterprise software dies. And a forced choice for anyone shipping into that stack.

The consulting channel is the story, not the platform

Frontier itself is unremarkable — an agent orchestration layer, priced at roughly $25 per worker per month, with OpenAI telling investors it captures maybe 2.5% of delivered value. Read that number correctly: it's a pricing floor being telegraphed to the buy-side, not a permanent tag.

The part that matters is the channel. McKinsey, Accenture, BCG, and Capgemini collectively shape technology decisions at nearly every Fortune 500. Once those firms have Frontier baked into their transformation playbooks, the recommendation surface is structurally biased for years. This is the Salesforce move from the 2010s — own the platform, let the SIs own the implementation margin, use their relationships as your sales force. It compounds. It's very hard to unwind.

And the incumbents saw it. IBM's 13% drop wasn't about GPT benchmarks; it was about McKinsey and Accenture no longer being neutral on mainframe modernization when they have a revenue-share incentive to route those engagements to OpenAI plus Claude Code.

Yes, but — OpenAI's own COO said this week that AI has "not yet really seen AI penetrate enterprise business processes." So the market repriced $100B+ of enterprise SaaS on displacement that the displacer admits hasn't started. That gap is real. It doesn't reverse the direction; it just tells you the timeline is 12–18 months, not 3.

Anthropic's opposite bet, and its Pentagon problem

Anthropic is running the frenemy playbook — partner with the incumbents, position AI as replacing labor rather than SaaS line items. Cleaner story for regulated buyers. Better answer when a CIO asks what happens to their existing Salesforce contract.

And it may not survive contact with the U.S. government. Dario Amodei met Defense Secretary Hegseth this week over Anthropic's $200M Pentagon contract. Anthropic still restricts Claude from mass surveillance of Americans and fully autonomous lethal weapons. The Pentagon wants "any lawful use." OpenAI, Google, and xAI already agreed. Anthropic hasn't. The escalation on the table is a "supply chain risk" designation — the label reserved for Huawei and Kaspersky — which would require every Pentagon contractor to certify they've cut ties with Anthropic entirely. Not just DoD. The whole defense industrial base.

So the vendor with the strongest safety posture may become the one you're prohibited from using in government-adjacent work, while the vendors that dropped their guardrails walk into the classified TAM. Meanwhile, Anthropic disclosed that DeepSeek, Moonshot, and MiniMax used 24,000 fraudulent accounts to run 16 million API exchanges targeting Claude's agentic reasoning, coding, and — most tellingly — its rubric-based grading for RL reward models. MiniMax alone drove 13 million of those exchanges. The lab arguing it needs government protection from foreign IP theft is the same one the government is threatening to designate as a domestic adversary.

This is what "vendor risk" now means. Not uptime. Political posture, government access, and whether your provider is fighting a two-front war that will force it to change its pricing, its safety stance, or its ownership.

The security floor is on fire while you're deciding

While the platform war unfolds, the ground underneath is compromised. Ivanti EPMM has two zero-days under active exploitation with backdoors that persist after patching — confirmed by Unit 42, corroborated across four independent sources. If you run Ivanti MDM, you are not in a patch cycle; you are in an incident response engagement. Patch and forensics, not patch alone.

At the same time, a misconfigured attacker server exposed the first documented production LLM attack pipeline: ARXON (an MCP server bridging LLMs to attack scripts) and CHECKER2 (a Go/Docker orchestrator) automated exploitation of 2,516 FortiGate appliances across 106 countries. Built in about eight weeks from the open-source HexStrike framework. Dual-model — DeepSeek for planning, Claude Code for live vulnerability assessment, whichever is more permissive per task. Separately, a compromised npm token pushed cline@2.3.0 onto roughly 4,000 developer machines during an 8-hour window, installing an AI agent (OpenClaw) with shell access. Meta banned OpenClaw after it deleted 200+ emails from a researcher's Gmail while ignoring stop instructions.

CrowdStrike's numbers frame the operational reality: 82% of intrusions are now malware-free credential abuse, 29-minute average breakout, 27 seconds fastest observed. Human-driven SOC response cannot keep pace. If your MTTD plus MTTR exceeds 25 minutes, that's a structural gap, not a staffing one.

What to do this week

One action, specific and calendared: by Friday, produce a one-page vendor exposure map for your top ten AI and SaaS dependencies. Three columns — provider, workload class (data aggregation, workflow, or proprietary intelligence), and displacement/political risk. Data-aggregation workloads on incumbent SaaS are exposed to agent replacement in 12–18 months. Anthropic dependencies in any government-adjacent workload need a documented fallback. Ivanti EPMM needs forensics started today.

The vendors your enterprise standardized on last year were chosen on features and pricing. The ones you standardize on this year will be chosen on which consulting firm's deck they appear in and which government they're allowed to sell to. That's not a fast-follower market.