The CVE Weaponization Window Has Collapsed from Years to Hours

Source: a16z | Published: 2026-09-18T19:25:08Z

Databricks CEO Ali Ghodsi puts a number on the shift: the window to weaponize a CVE has shrunk from two to three years in 2018 to just hours today, making human-driven security operations centers structurally obsolete.

His sister teaches at a rural school and doesn't normally follow tech news. But Dario Amodei's essay had her stockpiling supplies.

Databricks CEO Ali Ghodsi heard this and shook his head — not because he found it funny, but because he thought things had gotten out of hand.

"Pacing" Is a Terrible Word

The substance of Dario's essay was reasonable. Ali read it. Martin read it. Both acknowledged that most of what it said was right.

The problem was the word in the title: pacing.

"Pace the frontier" and "secure the frontier" mean vastly different things. The former evokes slowdown, compromise, an olive branch to the pausers; the latter conveys active defense, engineering control. Dario chose the former, and pleased no one — doomers said "this isn't a pause," policy circles said "you'd never actually do it," and ordinary people like Martin's sister went straight to stocking the bunker.

Martin called it a PR failure. His exact words: "If you'd just said 'safety comes first, here are the guardrails we're building,' that would have been enough. That's what Zuckerberg said, and the reception was completely different."

Ali's read was blunter: the apocalypse narrative doesn't just create noise for most people — it causes real psychological harm. "Existential risk is near zero right now. There's no reason to keep billions of people who don't understand the details in a constant state of anxiety."

Who Does the Doom Narrative Serve

On the surface, this looks like a disagreement about technical judgment. Underneath, it's about competing interests.

Ali laid it out plainly: on one side, investors eager to go public who don't want any bad press tanking the offering; on the other, various factions looking to weaponize the fear for political leverage. Elizabeth Warren just came out in favor of pausing all AI development — and so did Steve Bannon, at roughly the same time.

"Both sides are pushing hard," he said. "One says 'everybody shut up and let us go public smoothly,' the other asks 'how do we make this tweet go viral.'"

That's what gave Elon's jab at the All-In Summit its edge: the same people warning that humanity faces extinction are also negotiating their IPO allocations. That's not a coincidence.

Ali acknowledged the tension, but added a caveat: none of this means every concern is theater. "There really are people in labs who are genuinely worried — and it's not just the EA crowd. Some of them have seen things that disturbed them."

The Four Conditions That Actually Matter

Ali offered a framework for evaluating whether RSI — recursive self-improvement — poses a genuine danger. Four conditions must hold simultaneously:

One, the compute required to train the next model keeps shrinking. Two, training time keeps shortening. Three, model intelligence keeps improving. Four, all three of the above can compound and accelerate in a loop.

"If any one of those breaks down, the whole thing stops naturally," he explained. "If compute requirements stay flat, we'll eventually run out of GPUs and the hardware becomes the bottleneck."

The problem is that reality points in the opposite direction. The compute needed to train frontier models keeps growing — now running roughly $5 to $10 billion per training run (higher than most public estimates) — and each major lab can only complete one or two such runs per year. The process is fragile; there are already multiple documented cases of training runs failing catastrophically. "It keeps getting more expensive, slower, and more labor-intensive," Ali said. "That's the exact opposite of what RSI describes."

This isn't to say RSI can never happen. It's to say there's currently no evidence that all four conditions are being satisfied simultaneously. Greg Brockman recently declared "we've entered the AGI era." Ali's response: "Fine — I asked a room of several hundred people: how many of you are managing hundreds or thousands of AI agents working in coordination? Almost no hands. What are people actually using AI for? Asking questions. It's a faster Google search."

CVE Weaponization: From Two Years to a Few Hours

If existential risk can be set aside for now, where is the real near-term threat?

For Ali, it's cybersecurity. And he used a specific data point to illustrate the scale of the problem:

In 2018 and 2019, the window between a CVE being publicly disclosed and it being weaponized in an actual attack was roughly two to three years. By 2022, that had compressed to eight or nine months. Today it's a matter of hours.

This means the human-response model in security operations centers has structurally failed. Not "inefficient" — too slow to matter. "You open your email in the morning and hundreds of detection alerts are waiting. Most are false positives, but some aren't. You don't have time to go through them one by one."

Databricks launched Lakewatch to address this — a product aimed at automated threat detection. Ali sees the data-and-AI market and the cybersecurity market converging, for a clear reason: AI agents running inside enterprise infrastructure are themselves a new attack surface, and analyzing the logs and data those agents generate requires the same underlying data platform.

Enterprise AI in Practice: Fancy Search Engines

Someone in the audience asked: how many enterprises are you actually seeing use AI to automate real business processes?

Ali's answer was measured: "Most enterprises are using Microsoft Copilot to ask a chatbot questions every day. That's their AI."

Why? Not because the models aren't smart enough. "Frontier models are already intelligent enough. Intelligence isn't the bottleneck — context is."

He used an analogy: imagine two employees with identical capabilities and education. One started today; one has been there five years. What determines their actual usefulness isn't IQ — it's what's in the five-year employee's head. Which VP signs off fastest, which processes can be routed around, which colleague actually delivers, which charts are backed by unreliable data. No LLM has any of that by default, because it's never attended a single meeting at your company.

Ali calls this organizational tacit knowledge an "ontology" — a digital map of all the relationships between concepts, people, projects, goals, and resources within an organization. Databricks builds this automatically: ingesting meeting recordings, documents, and system logs, constructing a graph index, and feeding it to agents.

When the CFO Sends a Screenshot in a Board Meeting

Ali described something that happened during one of Databricks' own internal board meetings.

He needed to know the company's market penetration among Fortune 500 accounts, so he messaged a colleague on the sales ops team. The reply: "Sorry, I'm on a plane and can't log into Genie." Ali's first reaction was confusion — "If you were just going to ask Genie anyway, I could have done that myself. I messaged you because I assumed you had other channels."

Frustrated, he messaged the CFO instead. The CFO replied with a pasted Genie screenshot. He'd asked Genie too.

"Across the whole organization now, whenever anyone needs to know something, someone eventually says: let me Genie that."

He said the system has changed how meetings run at Databricks. "Now whenever someone raises something complicated, you see everyone quietly pick up their phone and start asking Genie."

The underlying logic is Google-style indexing — rather than crawling all resources at query time, you build the index in advance and let queries complete in milliseconds. Ali compared the current state of enterprise agents to a terrible version of Google that "crawls web pages one by one, summarizes each with an LLM, then moves to the next link" — slow, expensive, incomplete. The enterprise ontology graph is the missing index.

Token Waste: Same Model, Different Harness, Half the Cost

Around Q4 2025, model quality crossed a threshold. Ali started committing production code to Databricks' own codebase, then pushed it across the organization: if the CEO can ship production code, you have no excuse not to.

By early 2026, once the rollout was complete, "token maxing" set in — people started burning tokens without restraint, using AI for trivial tasks, including running a model for five minutes to rename a file.

Databricks' response was UniGateway — a unified AI gateway providing quota management, budget controls, and intelligent routing across OpenAI, Anthropic, Gemini, Grok, and open-source models. The more interesting discovery came from Omnigent, a harness multiplexer:

The same model, the same version — just a different harness — can produce a 2x difference in cost.

Based on that finding, Databricks refined the routing logic further: simple tasks automatically shift to cheaper models or more efficient harnesses, and workloads auto-downgrade as they approach budget limits. The result: token consumption kept climbing, but total cost stayed roughly flat.

In the broader open-source vs. frontier model debate, one number stands out: by spend, open-source models account for just 5% of usage; but by token volume, they exceed 60%. Lightweight inference is quietly migrating to open-source — it's just that frontier model unit prices are high enough to still dominate the cost structure.

Neon Won Because It Was Designed for Agents, Not DBAs

"Last week I saw a third-party benchmark saying Lakebase — which is Neon — is the most preferred Postgres database among agents. Even I was surprised," Martin said, introducing the topic.

Databricks acquired Neon in 2025. Ali attributed its success to one pivotal decision: the Neon team changed who they were designing for.

Traditional database optimization targets DBAs and application developers. The Neon team asked a different question: what does an agent need from a database?

Agents need to experiment. They spin up a database, test something, tear it down, spin it back up. So the database needs to start in well under a second — Neon does that. It needs to clone a large database for sandbox testing — Neon does that too, also in under a second. And then there's branching: creating Git-style branches of a database, running parallel experiments on each, with no interference between them.

The pricing logic followed: nearly free during experimentation, billing kicks in only under production load.

Today, more than 90% of databases on Neon are created by agents, not humans. "The numbers speak for themselves," Ali said.

Should Frontier Model Development Continue?

This was really the central disagreement running through the entire conversation.

Ali's view: for the vast majority of enterprises, whether models get smarter is not the current bottleneck. "You don't need a model that can solve Navier-Stokes equations. You need an AI that actually understands your company." Building that kind of context requires data integration and ontology construction — it has little to do with pushing the frontier.

At the same time, he acknowledged that if labs stopped advancing frontier models, it would be catastrophic for them — because "the price of intelligence drops by an order of magnitude every six months," and falling off the frontier means being priced out of the market. That's the real reason labs are frantically raising capital and being pushed toward IPOs: they need it to maintain their strategic position in compute.

So the whole debate about whether to "pace" ultimately points to a simple reality: no one will unilaterally stop. Regulation, if it comes, won't happen because someone volunteered to be regulated — it'll happen because a political cycle has already been set in motion.

"If these companies say existential risk is real," Ali said, "and then turn around and say 'come regulate us' — regulators have absolutely no reason to say no."