Thursday, and the theme writes itself: artificial intelligence is now sitting on both sides of the security table at once — and someone has to pay for the chairs. Today a new model ships exclusively to defenders, a research model proves the offense is real, and the company funding much of the buildout reminds investors what it costs. If you own the risk and the budget, this is your news cycle.
Google ships a model only defenders can use
The notable model release today isn't a general-purpose upgrade. Google introduced Gemini 4 Carbon, a model built for cybersecurity work and released first with a restricted rollout to cyber defenders. The framing is the story: Google released Gemini 4 Carbon with a focus on cyber defense rather than shipping it to everyone at once.
The operator's take: gating a frontier model to defenders is an admission that the same capability cuts both ways, and the vendor would rather not hand the offensive version to anyone with a credit card. For you, a defender-only model is useful only if you have the telemetry and the people to feed it — a smarter analyst does nothing for a team that still can't see its own attack surface. Buy the tool after you've bought the visibility, not before.
An unreleased Claude builds real exploit chains
If you wanted proof the offensive side is not theoretical, you got it. Anthropic's Frontier Red Team disclosed that an unreleased Claude research model built end-to-end exploit chains across five platforms, including a zero-click remote-root Linux chain (CVE-2026-98432, CVSS 9.8), a one-click Chrome escape combining four V8 and sandbox bugs patched in Chrome 156, a Safari same-origin break through WebKit's Gigacage, a locked-Pixel-9 bypass over USB, and a Windows QUIC crash. This is a model chaining individual bugs into working intrusions, not a chatbot describing one.
The operator's take: the hard, expensive part of an attack — stitching separate vulnerabilities into a reliable kill chain — is exactly the part a model just demonstrated it can do. That compresses the window between "patch released" and "patch weaponized," which means your patch SLA is now your real exposure window. Shorten it. The unglamorous controls win here: patch the edge fast, kill unused attack surface, and assume anything disclosed will be exploitable sooner than it used to be.
Alphabet shows the bill for the arms race
The money behind all of this surfaced in earnings. Alphabet reported Q3 2026 revenue of $127.6B, up 25% year over year, with Google Cloud revenue up 103% to $30.7B and capex of $51.4B in the quarter; CEO Sundar Pichai said first-party model APIs now process nearly 32 billion tokens per minute and the Gemini app has reached 1 billion users. The market still flinched: Alphabet shares fell about 7%, and the company has signaled annual capex will top $200 billion before exceeding $300 billion in 2027.
The operator's take: a cloud business doubling while the stock dips on spending is the whole AI economy in one chart — the demand is real, and so is the capital bonfire required to serve it. For you, the downstream signal is cost: that $200-billion-and-climbing buildout gets amortized into the per-token and per-seat prices you'll pay over the next few years. Lock in what you can, keep your model and compute choices portable, and don't architect around today's promotional pricing. The vendors are spending like the music never stops; price your contracts like it might.
Also on my radar
- The chip money keeps consolidating. AMD announced a $1 billion investment in Meta's equity offering to deepen their AI-infrastructure partnership, with Meta already running millions of EPYC processors and Instinct GPUs in production — the hyperscalers and chipmakers are now tying their balance sheets together, which is worth watching if you depend on either staying independent.
- Supply-chain defense gets a default. GitHub launched an npm package firewall that can block brand-new malicious package versions before they install — if you ship JavaScript, this is a cheap control to turn on before the next poisoned-dependency incident, not after.
- A nation-state zero-day lands in the catalog. CISA added a Broadcom VMware ESX flaw to its Known Exploited Vulnerabilities list after CrowdStrike tied it to a China-nexus actor, Warp Panda, targeting the U.S. legal sector — KEV additions are your free, prioritized patch queue; if ESX is in your estate, it just jumped the line.
The throughline today is that AI has stopped being a tool you point at security and become a participant in it — defending from one chair, attacking from another, and sending the invoice from a third. Google's defender-only release and Anthropic's exploit-chain disclosure are the same admission from opposite ends: capability is now symmetric, and whoever operationalizes it first wins the exchange. Alphabet's numbers are the reminder that none of it is free. The operator's job isn't to pick a side of the table — it's to make sure your defenses, your patch cadence, and your contracts all assume the other chairs are already occupied. That's the Signal for today.
Paul Sapio is the CIO of Mikhail Education and a full-stack AI engineer. Open to contract work in security, networking, AI, and SaaS development — reach out.