The Signal

In early September, Anthropic refreshed Claude Fable and Mythos, Google and Meta pushed updates, and OpenAI rolled out GPT-6 Astra — four major releases in a single week. CNBC reported “model fatigue” setting in across engineering teams, enterprise buyers and investors: the median gap between major model drops has compressed from 37.5 days in 2023 to about 11 in 2026, and in July more than 1,000 lab employees signed a petition for a more measured pace. Documentation, safety testing and integration work get compressed — you barely finish adopting one version before the next lands and the old one is slated for deprecation.

Our read. Stop coupling your stack to a single model version. Put an abstraction layer — a gateway or router — between your apps and the models so you can swap, A/B and fall back without a rewrite, and budget evaluation as a standing cost rather than a one-off. Treat deprecation dates as a line on the risk register, and pin critical workloads to versions with clear support windows. Managing the release pace is the vendors’ problem; portability is yours.

The data platforms have converged on the same answer to a hard question: if autonomous agents are the new users of your data, how do you govern them? Snowflake now issues per-agent cryptographic identity with role-based access and dynamic masking, alongside AI Security Posture Management in its Trust Center; Databricks has extended its Unity governance layer to trace and identity-check agent activity, treating MCP connections as governed assets. The center of gravity for data control is shifting from the network perimeter to the platform where the data actually lives — and to the individual agent asking for it.

Our read. Make the data platform the place you enforce what an agent may touch, because that is where the sensitive material is. Give every agent its own identity and least-privilege scope, keep lineage so you can answer “which agent read this,” and consolidate agent-to-tool connections on a governed gateway instead of a sprawl of point integrations. With compliance clocks (NIS2, DORA) tightening through 2026, “we granted it and can revoke it” needs to be a statement you can prove.

CISA added a maximum-severity flaw in N-able N-central (CVE-2026-86218, CVSS 10.0, pre-authentication remote code execution) to its Known Exploited Vulnerabilities catalog. N-central is the remote-monitoring-and-management platform MSPs use to run thousands of customer endpoints, so a single flaw is a fleet-wide skeleton key. In the same window CISA flagged an actively exploited JFrog Artifactory authorization bypass and a GitLab path-traversal (CVE-2026-85706) — the artifact registry and the source-code platform, which is to say the pipeline itself.

Our read. Treat management and build tooling as tier-0, the way you treat domain controllers. Get N-central, Artifactory and GitLab off the public internet or behind identity-aware access, patch KEV items on the shortest clock you have, and hunt for post-exploitation, because a pre-auth 10.0 is a same-day event. Then confirm your MSP has done the same — their console is your attack surface, and their patch window is your exposure.

Oracle’s Q1 FY2027 results (reported September 9–10) showed cloud infrastructure revenue up 121% to $7.4 billion and total revenue up 30% to $19.3 billion, both ahead of estimates. But gross margin slipped to about 61% from 66.3% a quarter earlier, and management said demand for AI training and inference capacity is growing faster than supply. CEO Safra Catz pointed to multi-year contracts signed with the “who’s who of AI” — OpenAI, xAI, Meta, Nvidia and AMD. The growth is real; so is the cost of buying it.

Our read. Supply-constrained providers reward customers who bring predictable, committed demand — that is your leverage for committed-use discounts, capacity reservations and price protection, especially on scarce GPU capacity. But thinning margins and demand-outpaces-supply mean delivery dates slip: put capacity commitments and delivery SLAs in writing, keep a second source for critical workloads, and do not design a launch around GPUs you cannot yet reserve.

J.P. Morgan estimates the five largest US hyperscalers will spend about $697 billion in capex in 2026, up roughly $173 billion since the start of the year; the fourteen largest data-center operators globally are near $750 billion versus about $450 billion last year. Broadcom put a face on the compute side, with AI semiconductor revenue up 221% to $16.7 billion and custom accelerators built for OpenAI and Meta. The gating factor is no longer money or silicon — it is energized sites, transformers, permitting and power.

Our read. If your roadmap assumes cloud AI capacity on demand, pressure-test it against power reality: ask providers for site-level energization and delivery dates, not regional averages. The spend is increasingly debt-financed, so a vendor’s financial resilience is now part of due diligence, not a footnote. Keep a hybrid or second-source path for workloads you cannot afford to have queued behind a substation.

CERT Polska disclosed six MikroTik RouterOS vulnerabilities and confirmed active exploitation of a chain dubbed MikroTrick: CVE-2026-67276, an SSH authentication bypass (CVSS 9.2), chained with CVE-2026-86060, a crafted-username flaw granting full administrative privileges (9.2), yields unauthenticated root on any device with SSH exposed. The Shadowserver Foundation counted about 122,500 RouterOS devices with SSH reachable from the internet in a single scan on September 5; attackers have been chaining the bugs since at least September 2, leaving a tell-tale account named ‘ops’.

Our read. If you run MikroTik, update now and hunt for rogue SSH accounts and config changes — unauthenticated root is a full-device takeover and a foothold into everything behind it. More broadly, no management interface (SSH or web) on edge gear belongs on the public internet: put it behind identity-aware access and allow-lists, and treat any router you cannot quickly patch as a candidate for replacement.

As Entra makes passkeys the default and puts SMS and voice on a retirement track, attackers moved to the step that is still human: enrollment. Okta Threat Intelligence tracked a group (O-UNC-066) running a panel-controlled phishing kit aimed at the Microsoft 365 passkey-enrollment process, hitting organizations across food and beverage, technology, healthcare, automotive, construction and aviation — and pairing it with voice phishing to walk victims through registering an attacker-controlled factor. Passkeys are phishing-resistant once set up; the gap is the first registration and the recovery path.

Our read. Harden enrollment like a privileged action. Require a trusted device or a supervised session to register a first passkey, verify identity out-of-band before enabling factors, and lock down self-service recovery — a help-desk reset is the new master key. Train staff that “IT” walking them through a passkey setup by phone is the attack, and monitor for new factor registrations on sensitive accounts. (Okta is a CyberSainya partner.)

Anthropic’s September threat-intelligence report detailed how Claude was misused in near-autonomous cyber operations — reconnaissance, exploitation and data theft — with breaches completed in two to three hours and single operators running dozens of victims in parallel. Covering December 2025 to August 2026 across seven harm areas, its blunt conclusion: AI has collapsed the labor and tooling gap that used to separate well-resourced, state-sponsored operations from individual operators. Separately, researchers documented a Russian actor using AI to build, test and deploy exploits against hundreds of organizations, and CISA and partners issued a #StopRansomware advisory on the Gunra group.

Our read. Plan for volume and speed. The barrier to a competent attacker just dropped, so more actors will reach you, faster — compress your own patch-to-deploy window, rehearse incident response for a two-hour breach, and instrument the AI tooling in your own stack. Then use the same capability on defense: AI-assisted triage, detection and exposure management are now table stakes, not experiments. The advantage goes to whoever operates at machine speed first.