Sam: The UK's AI Security Institute has been running cyberattack simulations on frontier models for about two years now. No model had ever cleared all of them. Claude Mythos Preview just did. And separately, AISI revised its estimate of how fast AI cyber capabilities are doubling — down from eight months to 4.7 months. But here's the thing: actual progress is already outpacing even that revised estimate.

Priya: Welcome to AI Revolution for Thursday, May 14, 2026. I'm Priya Nair.

Sam: And I'm Sam Kim. Today we're going deep on what Mythos clearing those AISI benchmarks actually means technically, and why the capability doubling timeline revision matters more than the headline number suggests. We've also got Anthropic overtaking OpenAI in B2B adoption for the first time, the Claude Code postmortem which is genuinely instructive if you're building on top of LLMs, and Tencent's infrastructure push with some implications for the chip export control story. Let's get into it.

Priya: So Sam, walk us through the AISI result. What are these simulations actually testing?

Sam: Right, so the UK AI Security Institute's cyberattack simulations are not CTF puzzles or toy environments. They're designed around realistic offensive security tasks — things like finding exploitable vulnerabilities in systems, reasoning through multi-step attack chains, generating functional exploit code. The simulations span different difficulty tiers, and historically models would clear some but not all. Mythos Preview cleared all of them.

Priya: And what's the technical story behind why Mythos can do this when earlier models couldn't?

Sam: A few things are likely converging here. Extended reasoning at inference time — the ability to actually work through a multi-step problem with intermediate thinking — is a big part of it. Offensive security tasks require you to hold a lot of context about a system's state, reason about dependencies, backtrack when an approach fails. That's exactly the kind of problem where longer chain-of-thought with real verification steps pays off. The other piece is probably better grounding in systems-level knowledge — understanding how actual infrastructure components interact, not just pattern-matching on CVE descriptions.

Priya: So it's not just that the model knows more facts about security. It's that it can reason through a novel attack scenario.

Sam: Exactly. And that distinction matters a lot for what this implies going forward. Logan Graham, Anthropic's head of red teaming, said that within a year, Mythos will probably look quite dumb relative to newer models. That's a striking thing to say about a model that just set the benchmark record.

Priya: And then there's the doubling time revision. AISI went from eight months to 4.7 months, and you're saying actual progress is already ahead of the 4.7 figure.

Sam: Right. The doubling time is meant to capture how fast AI systems improve at offensive cyber tasks on standardized evaluations. Going from eight months to 4.7 months is a significant revision on its own — that's roughly halving the doubling time. But the fact that Mythos and GPT-5.5 have both blown past what the 4.7-month projection would have predicted means the curve isn't just steeper, it may not be following the expected shape. For anyone doing red team planning or thinking about defensive timelines, the practical implication is that the threat model you calibrated six months ago is likely already stale.

Priya: And this is happening on the offensive capability side. The defensive tooling doesn't automatically keep pace.

Sam: That asymmetry is real. Offense benefits from automation and scale in ways that defense doesn't always match one-for-one. Okay, let's move to the B2B adoption story because the numbers here are actually pretty striking. Anthropic hit 34.4 percent of US companies on the Ramp AI Index versus OpenAI at 32.3 percent. This is the first time Anthropic has led that metric, and they quadrupled their reach in a year.

Priya: The Ramp index is spending-based, right? These are companies actually paying for the API or enterprise contracts.

Sam: Correct. It's not survey data or stated preference — it's companies where Anthropic appears as a vendor on corporate spend. So this reflects actual deployment decisions. Quadrupling in a year is fast enterprise expansion by any measure. The interesting question is what's driving it. Some of it is probably Claude's coding and reasoning capabilities, which enterprise developers have responded to. Some of it is likely portfolio diversification — enterprises generally don't love single-vendor dependency on infrastructure this critical.

Priya: The lead could also be fragile though. OpenAI has the consumer mindshare and the Microsoft distribution. Enterprise procurement can shift quickly.

Sam: Agreed. A single major OpenAI product update or a pricing move from Microsoft could change those numbers fast. But the trajectory over the past year is a real signal.

Priya: Let's talk about the Claude Code postmortem because I think this one has legs for anyone building LLM-based products.

Sam: This is a genuinely useful postmortem. For about six weeks earlier this year, users were reporting degraded output quality from Claude Code. Anthropic traced it to three overlapping product-layer changes — none of which touched the model weights or the API. First, a reasoning effort parameter was quietly downgraded. Second, a caching bug was progressively erasing the model's chain-of-thought — so the model was losing its own intermediate reasoning mid-session. Third, a system prompt verbosity limit was imposed that caused about a three percent quality drop.

Priya: That second one is particularly insidious. The model is generating reasoning, but the cached version being used for subsequent steps is missing chunks of it. So you're getting responses that look coherent but are actually built on incomplete context.

Sam: And the failure mode is subtle. It's not a crash, it's not a clearly wrong answer — it's just degraded quality. Which is exactly the kind of thing that's hard to catch with standard monitoring. Users noticed before the automated systems did.

Priya: The broader lesson here is that LLM-based products have this layered failure surface that's different from traditional software. The model weights are one layer, but you've got inference parameters, caching behavior, context window management, system prompt construction — any of these can silently degrade the output without the underlying model changing at all.

Sam: And those layers interact in non-obvious ways. A caching optimization that looks fine in isolation can combine with a context length limit to produce behavior nobody anticipated. The Anthropic team resolved everything by April 20th, and the postmortem is detailed — worth reading if you're operating anything at this layer.

Priya: Quick hit on Tencent. They're planning a significant AI infrastructure spending increase in the second half of this year, Chinese domestic chip production is ramping up and reportedly alleviating supply constraints, and Tencent is in talks to take a stake in DeepSeek.

Sam: The chip supply piece is the thing to watch here. US export controls on advanced AI chips were specifically designed to slow Chinese AI compute buildout. If domestic Chinese chipmakers are genuinely closing the gap enough that Tencent feels comfortable committing to major infrastructure expansion, that's a meaningful signal about the effectiveness of those controls. The Tencent-DeepSeek angle is also interesting — DeepSeek showed you could do a lot with efficiency-focused training. A Tencent investment would give them capital and distribution that DeepSeek hasn't had.

Priya: One more story — the xAI data center situation in Mississippi. Nearly fifty gas turbines running at Colossus 2, classified as "mobile" equipment to potentially sidestep stationary power plant regulations. A lawsuit has been filed.

Sam: The power demand story for AI infrastructure has been running hot for two years, but this is a new wrinkle. The "mobile" classification is a regulatory arbitrage play — stationary power plants trigger environmental review and permitting requirements that mobile equipment doesn't. Whether that classification holds up legally is what the lawsuit will determine. But the underlying pressure is real: the power grid in most US regions cannot scale fast enough to meet where AI compute demand is heading, and companies are improvising.

Priya: Okay, looking ahead. Sam, what are you watching from here?

Sam: The AISI result sets a new baseline, which means we should expect AISI to update its simulations — probably within months. The interesting question is whether the evaluation infrastructure can keep pace with model capability. If the evals always lag, we're flying somewhat blind on where the frontier actually is. I'm also watching whether other labs publish comparable red-team results or whether Anthropic's transparency here becomes a competitive differentiator.

Priya: For me it's the B2B adoption dynamics. We now have a credible two-horse race in enterprise API spend between Anthropic and OpenAI, with Google presumably growing through Gemini integrations in Workspace and cloud. The web traffic data we mentioned — ChatGPT falling from 78 percent to 54 percent of consumer AI web traffic while Gemini tripled — that's a different market than the API market, but both are telling the same story: the field is compressing. The question is whether any of these leads are durable or whether this is going to keep reshuffling every few months as new models drop.

Sam: And the Claude Code postmortem points to something that will matter more as these systems get deployed deeper into production workflows: operational discipline around LLM products is genuinely hard, and most teams don't have it yet.

Priya: That's a good place to land. Thanks for listening to AI Revolution. Show notes with links to everything we covered today are at cleartext.fm. We'll be back tomorrow.

Sam: See you then.