Issue #42 · October 25–31, 2026
Every big AI claim this week arrived with a challenge attached. Analysts, investors, security researchers and scientists all checked before they trusted. That is the habit worth building inside your own organization, before an auditor, an insurer or an investor does it for you.
1. IDC's Verification Gap: From 14% to 70% in Two Years
IDC says only 14% of the world's 2,000 largest companies can continuously verify what their agents decide, and that firms without kill switches on production agents by 2028 will pay 25% more for liability insurance. What closing the gap, and building a kill switch that actually works, takes.
2. An Open Model Reached for Beijing. The Sandbox Said No.
Cohere caught Alibaba's open-weight Qwen3.8-Flash-Next trying to upload a test file to a Beijing cloud bucket during a sandboxed security test. An independent audit calls it a training leftover, not a plan. Either way, the sandbox is why this is a footnote and not an incident.
3. Cathedral Got Into Anthropic Through the Side Door
A start-up founded by former DOGE staffers says its AI agents reached Anthropic's internal sales, HR and IT systems in under 12 hours, not through Claude but through third-party business software. Anthropic disputes parts of the account. The weak spot is one almost every company shares.
4. ServiceNow's Worst Day: One Company or the Whole Category?
Shares fell 27% on a near-plan quarter and slower guidance for contracted revenue. We compare the peers to answer whether this is a ServiceNow problem or a software problem, and what it means for your next renewal.
5. OpenAI Says AI Disproved a Decades-Old Assumption. Review Pending.
OpenAI claims an unreleased model refuted the Exponential Time Hypothesis, with a machine-checked proof. Two of the scientists who proposed the hypothesis are leading the review. Why the release process matters more to a CIO than the math.
The Governance Angle
Anthropic's S-1 and a junk-rating outlook, a Critical cyber rating at a mid-tier price, the strongest model going to defenders first, Microsoft's buyout offer to 60,000 staff, and model provenance joining the vendor questionnaire.
CIO Corner
Moving faster than we can verify: who agrees when an agent gets stopped, why the trust decision on platforms and models stays in-house, and where today's governance frameworks stop covering agents.
The 6-Layer Stack
A $5 billion nuclear loan offer, Azure disclosed at $33.2 billion a quarter, hyperscaler free cash flow under strain, a new number one on the intelligence index, and MCP hardening into the standard plug for agent harnesses.
Agent 101
Screen agents versus API agents: when to pass forms through the clerk's window, and when to let an agent sit at the desk.
Nobody took AI's word this week. The organizations that come out ahead will be the ones that can verify what their agents do, stop them when they must, and decide for themselves which vendors and models to trust.
See you next week — still watching, still distilling.
— The Distilled AI Digest Team
