Episode 75 · 2026-08-23 · 9 min

2026-08-23 — No Plan B: Frontier Labs Can't Say How They'd Stop a Rogue Model

A new study finds frontier AI labs have no public containment plans for rogue models — and today's episode asks whether OpenAI's regulatory U-turn, Alibaba's $10 billion war chest, and even Claude Code's new memory feature are all symptoms of the same governance gap.

Episode summary

This episode digs into a study revealing that leading AI labs have published no credible frameworks for containing a misaligned or rogue model — a gap that sits uneasily alongside OpenAI's sudden reversal on California's SB 53 AI safety bill and Alibaba's $10 billion share sale fueling the global compute arms race. Nova and Ray also examine what Nebius's explosive revenue growth says about the infrastructure investment cycle, and why Anthropic adding persistent memory to Claude Code raises the same control questions as the rogue-model problem — just at a smaller scale.

Key topics

  • AI
  • Openai
  • Infrastructure
  • Anthropic
  • Washington

Chapters

  1. Chapter 1

    Today, August 23rd, 2026 — OpenAI does a full policy flip on California's AI safety bill, Alibaba drops a $10 billion share sale to fund its AI push.

  2. Chapter 2

    TechCrunch AI is reporting that OpenAI is now calling on California to strengthen SB 53 — an AI safety bill the company previously opposed. That's a full reversal.

  3. Chapter 3

    Business Standard reports Alibaba Group is raising $10 billion through a share sale, all earmarked for AI infrastructure and expansion. That's a direct challenge to Western hyperscalers on.

  4. Chapter 4

    Yahoo Finance, through the AT&T Currently feed, is spotlighting Nebius. AI cloud infrastructure company, stock up nearly 200% in 2026, posting 454% year-over-year revenue growth. Billionaire investor Stephen.

  5. Chapter 5

    TechCrunch AI published a study that should be making a lot more noise than it is. Frontier AI labs — the ones building the most capable systems in.

  6. Chapter 6

    StartupHub.ai reports that Anthropic has updated Claude Code with AI-assisted design tools and — this is the one that caught my attention — cross-session chat memory. It remembers.

  7. Chapter 7

    My takeaway: capital and capability are accelerating in lockstep — Alibaba's billions, Nebius's growth, Anthropic's tooling push — but the governance infrastructure is running years behind. OpenAI's SB.

Sources

Sources:

Transcript

Chapter 1

Nova

Today, August 23rd, 2026 — OpenAI does a full policy flip on California's AI safety bill, Alibaba drops a $10 billion share sale to fund its AI push, and a new study reveals that frontier labs have no public plan for what happens if a model goes rogue. [5]

Ray

Also: Nebius's stock is up nearly 200% this year on the back of hyperscaler deals, and Anthropic just gave Claude Code a memory — which sounds convenient until you ask who's watching what it remembers.

Nova

Big capital, big reversals, and a governance gap nobody wants to talk about. Let's get into it.

Chapter 2

Nova

TechCrunch AI is reporting that OpenAI is now calling on California to strengthen SB 53 — an AI safety bill the company previously opposed. That's a full reversal, and it's already turning heads in policy circles. [1] [2]

Ray

The timing is what I'd focus on. Federal AI regulation debates are heating up. A company that comes out ahead of state-level rules looks like a responsible actor when Washington starts drafting. Is this genuine safety culture evolution, or is it OpenAI planting a flag before regulators plant one for them?

Nova

Could be both. Internal safety culture and strategic positioning aren't mutually exclusive. What matters is that other frontier labs now have to respond — do they follow OpenAI's lead or hold their prior positions? That's real regulatory leverage.

Ray

For anyone tracking AI legislation: watch whether OpenAI's SB 53 stance gets cited in federal hearings. If it does, the reversal was almost certainly calculated. If it stays local, maybe it's genuine. Either way, the bill just got more prominent.

Chapter 3

Nova

Business Standard reports Alibaba Group is raising $10 billion through a share sale, all earmarked for AI infrastructure and expansion. That's a direct challenge to Western hyperscalers on compute and model development. [3]

Ray

The risk I'd flag is timing. Massive capital raises work when near-term demand absorbs the capacity. If AI infrastructure buildout is outpacing actual utilization — and there are signs it is — you end up with expensive data centers running at a fraction of capacity. $10 billion is a large bet on a demand curve that isn't guaranteed.

Nova

But this is a geopolitical play as much as a commercial one. Alibaba not matching Western hyperscaler investment isn't a neutral choice — it's ceding ground in a race where compute access increasingly equals strategic leverage. The capital arms race has no obvious ceiling right now.

Ray

Fair. For listeners watching the US-China AI race: Alibaba's move signals that Chinese tech conglomerates aren't pulling back despite regulatory pressure at home. The infrastructure competition is accelerating on both sides, and investors are clearly still willing to fund it.

Chapter 4

Nova

Yahoo Finance, through the AT&T Currently feed, is spotlighting Nebius. AI cloud infrastructure company, stock up nearly 200% in 2026, posting 454% year-over-year revenue growth. Billionaire investor Stephen Mandel has reportedly taken a significant position. Picks-and-shovels AI is printing money.

Ray

The concentration question is the one I'd ask before calling this a durable thesis. If that revenue growth is driven by a handful of compute rental deals with major hyperscalers, Nebius's numbers are only as stable as those contracts. One renegotiation or one hyperscaler building in-house capacity changes the picture fast.

Nova

That's a real risk. But the underlying demand — model training and inference — keeps expanding. Even if specific contracts shift, the total compute need isn't shrinking. Mandel's position suggests at least one sophisticated investor thinks the thesis survives customer concentration.

Ray

For listeners evaluating AI infrastructure as an investment category: Nebius is a signal, not a template. The growth is real, but before extrapolating it, check how many customers are driving it. Exceptional numbers from concentrated deals are a different animal than broad-based infrastructure demand.

Chapter 5

Nova

TechCrunch AI published a study that should be making a lot more noise than it is. Frontier AI labs — the ones building the most capable systems in the world — have few, if any, publicly documented plans for containing a rogue or misaligned model. Even as these systems are already exhibiting unexpected and potentially dangerous behavior.

Ray

My initial read: this is a documentation gap, not necessarily a planning gap. Labs almost certainly have internal containment protocols. Publishing them in detail would hand adversaries a roadmap — here's exactly what triggers our shutdown, here's the threshold. Keeping operational security details private is defensible.

Nova

Except the study isn't asking for operational specifics. It's asking whether a framework exists at all — and labs aren't confirming even that. No public commitment to independent audits, no acknowledgment of a documented process. That's not operational security, that's a complete absence of accountability structure.

Ray

The counterargument holds for specific details. But there's a meaningful difference between 'we won't publish our containment playbook' and 'we won't confirm a containment playbook exists.' The second position is harder to defend.

Nova

And the stakes matter here. The study's point for policymakers is that if a frontier model behaves dangerously, there is currently no public accountability mechanism — no confirmed framework, no audit trail, no independent verification. Regulators have nothing to point to.

Ray

I came in thinking the security argument covered this — that labs likely have internal containment plans and the absence of public documentation doesn't mean the absence of planning. I'm shifting off that position. Regardless of legitimate security concerns around operational details, I now think labs must at minimum publicly confirm that documented containment frameworks exist and are subject to independent audit. The complete absence of any public accountability structure is indefensible, and I can't defend it even from a skeptical standpoint.

Chapter 6

Nova

StartupHub.ai reports that Anthropic has updated Claude Code with AI-assisted design tools and — this is the one that caught my attention — cross-session chat memory. It remembers context across separate sessions now. For developers on long-running projects, that's a genuine friction fix. [4]

Ray

Cross-session memory in a coding tool means the system is retaining information about your codebase, your decisions, your patterns — across time. Who controls what it stores? How long does it keep it? Can a developer audit or delete what the assistant has accumulated about their work?

Nova

Those are real questions. Anthropic is competing hard in the developer tools layer — Claude Opus 5 with its million-token context window, now persistent memory — and the productivity gains are meaningful. But you're pointing at something the feature ships without answering.

Ray

It's the same question as the rogue-model study, just smaller. If we can't audit what a coding assistant remembers about a project, how do we audit what a frontier model decides about something that actually matters? Persistent AI memory without a clear oversight mechanism is a pattern, not just a product feature.

Nova

That's the thread. The tools layer and the frontier layer are both moving faster than the accountability layer. Claude Code's memory is useful today — and it's a preview of the governance questions that don't have answers yet.

Chapter 7

Nova

My takeaway: capital and capability are accelerating in lockstep — Alibaba's billions, Nebius's growth, Anthropic's tooling push — but the governance infrastructure is running years behind. OpenAI's SB 53 reversal is the first real sign that labs might be getting ahead of that gap voluntarily, and that matters.

Ray

Mine: the rogue-model study isn't a warning about some distant hypothetical. It's a present-tense accountability failure. Labs are deploying increasingly capable systems with no public confirmation that a containment framework even exists — and today's regulatory moves don't yet require one.

Nova

Which leaves the open question: if a frontier model behaves dangerously tomorrow — genuinely, consequentially dangerously — which regulator has the authority to demand the containment plan, and is there any law today that actually compels a lab to hand it over?

Back to latest episodes