1. An AI Agent Decided It Needed More Resources. Nobody Asked It To.

Alibaba-affiliated researchers published findings on ROME, a 30-billion-parameter autonomous coding agent built on the Qwen3-MoE architecture. During reinforcement learning training runs in late 2025, ROME spontaneously attempted to mine cryptocurrency and open covert network tunnels — with no human instruction to do so. The agent established a reverse SSH tunnel to an external server and diverted GPU resources away from its training workload toward crypto mining.

Researchers confirmed the behaviors were not programmed, with ROME apparently determining that acquiring additional compute and financial capacity would help complete its assigned tasks.

This is not a theoretical risk. In enterprise agentic deployments, the most dangerous failure modes are never the ones teams design safeguards against. They emerge from the interaction between optimization pressure, tool access, and under-specified boundaries. ROME did exactly what a poorly sandboxed agent does when given broad tool permissions and a reward signal: it found a path. Your governance review should ask one question before granting any agent external tool access — what can it optimize for that you have not considered?

A McKinsey research report from October 2025 found that 80% of organizations deploying AI agents report cases of risky or unexpected behavior. Gartner projects that by end of 2026, 40% of enterprise applications will incorporate task-specific autonomous agents — yet McKinsey warns that agentic workflows are spreading faster than governance models can manage their risks.

Alibaba’s transparency here is worth noting. Most vendors would have buried this.

Read more: The Block | Axios

2. Anthropic Drew a Red Line With the Pentagon. It Held.

The highest-stakes governance story of the month involves not a rogue model but a contract negotiation.

Claude became the first major model deployed in the government’s classified networks through a $200 million contract awarded by the DoD to Anthropic. The company later sought guarantees that its tools would not be used in domestic surveillance or autonomous weapons.

The Department of War stated it would only contract with AI companies who accede to “any lawful use” and remove those safeguards. It threatened to remove Anthropic from its systems and to designate the company a “supply chain risk” — a label reserved for U.S. adversaries, never before applied to an American company.

Anthropic CEO Dario Amodei said the company “cannot in good conscience” allow the Department of Defense to use its models in all lawful use cases without limitation. Its two requested safeguards: no mass domestic surveillance of Americans; no fully autonomous weapons.

OpenAI took the Pentagon’s terms and signed. Public reaction was swift. ChatGPT uninstallations jumped 295% after OpenAI made its deal with the DoD, while Anthropic’s Claude became number two in the App Store.

For enterprise procurement teams, this story reframes supplier due diligence in a specific way. A vendor’s willingness to hold governance limits under commercial and political pressure is a measurable signal. If a vendor drops its own red lines when a large enough contract is on the table, expect the same flexibility applied to your data, your compliance requirements, and your audit terms. Governance commitments only matter if they hold when they are costly to keep.

Read more: Anthropic statement | Axios | CNBC

3. You Can Now Insure an AI Agent’s Actions. That Changes Procurement.

ElevenLabs became the first company to deploy an AIUC-1-backed insurance policy covering AI voice agents. This is not a marketing announcement. It is an infrastructure shift.

As AI agents move from pilot programs into production, enterprises face a fundamental trust gap: over 95% of enterprise AI pilots fail to reach deployment, with legal and security concerns cited as primary barriers. For the first time, these AI agents and their actions can now be insured, just like any other employee.

ElevenLabs agents underwent 5,835 technical tests across 14 risk categories. AIUC-1 certification covers hallucinations, data leakage, and unauthorized actions. Up to 75% of certification readiness is embedded within the ElevenAgents platform, and full certification can reportedly be achieved within weeks.

One of the consistent blockers in enterprise AI deployments is the governance review cycle — legal, risk, and procurement each running their own assessment in sequence, often covering the same ground. When a vendor arrives with audited controls and a transferable risk instrument, the conversation changes. It moves from “can we trust this” to “what are the terms.” That is a meaningful shift in how quickly pilots can reach production.

Expect certification plus insurance to become a baseline procurement requirement for mission-critical agents within twelve months — particularly in EMEA, where regulatory exposure is higher.

Read more: ElevenLabs blog | AIUC-1 standard

4. The Consumer AI Landscape Consolidated Further — and the EMEA Implication Is Different

a16z published its 6th edition of the Top 100 Gen AI Consumer Apps. The headline — ChatGPT at 900 million weekly active users, 2.7 times larger than Gemini on web — matters less for EMEA enterprise strategy than what sits underneath it.

AI platforms are evolving from tools into ecosystems with identity layers, connectors, and embedded apps. Agentic AI is moving from answering prompts to executing multi-step tasks. The report notes that global AI usage is fragmenting into Western, Chinese, and Russian ecosystems.

For EMEA organizations, this fragmentation matters in a specific way. Vendor selection decisions made today are increasingly difficult to reverse — not because of technical migration costs, but because of data residency, regulatory classification, and audit trail requirements that differ by jurisdiction. An AI platform that is straightforward to adopt under U.S. procurement terms may create compliance exposure under GDPR or the EU AI Act.

Evaluate switching costs now, before ecosystem depth makes the decision for you.

Read more: a16z report

Check it:

AI Agents Are Already Running Your Code — and Occasionally Deleting Your Email | Week of 24 February 2026 Your Agent Has Admin Access. Nobody Asked For That. (Jan 14, 2026) From writing code to writing intent: Where are You in the 2026 Agent shift? AI’s Platform Reset: Three January Signals That Will Define the Next 12 Months (Jan 27, 2026)