I post this now on its own so that we have a place to link to, to explain the situation.
Headline News
WSJ Headline (Obvious Nonsense): China Has Matched Anthropic in Cybersecurity, Resetting AI Race.
That. Did. Not. Happen.
The post even claims, explicitly, that Claude Opus 4.8 similarly ‘matches’ Claude Mythos, a claim which is even more obviously false.
Shame upon the Wall Street Journal. I fear Gell-Mann Amnesia. If they can get something as important as this so completely wrong, what about everything else?
I am skipping over the parts that involve accurate reporting, or minor quibbles.
It seems important to focus on clearly debunking the central false claims.
Alas, the mistakes made here very much rhyme with mistakes being made throughout all this by the White House, and that get latched onto by certain bad actors, who have played a large part in leaving us unprepared for the Mythos Moment.
It is important to understand what makes Mythos special. This is not it.
What Makes Mythos Special
What makes mythos special is not that only the chosen one can identify any given vulnerability in code.
What makes Mythos special is that it can identify vulnerabilities autonomously, at scale, without being pointed at them, and can then autonomously string together a variety of seemingly unrelated vulnerabilities into full working exploits.
This is the thing that GPT-5.6 Sol cannot do, and that Opus 4.8 and GPT-5.5 cannot do, and that GLM-5.2 also cannot do, at anything like the same level of difficulty. It is also a thing that Fable cannot do, not with any known prompting strategy, whether or not you call this a ‘jailbreak.’
Mythos cannot do anything that you could not eventually do with a less capable model, in the same sense that an infinite number of monkeys can write Shakespeare.
The public would net benefit if both Fable and Sol were generally available in their current forms as soon as possible, including on the question of cybersecurity. Whereas releasing Mythos to the public at this time would in expectation be a serious error.
Going Over The Detailed Claims
Robert McMillan, Raffaele Huang and Amrith Ramkumar (WSJ, being super misleading): Chinese artificial-intelligence systems have matched the performance of Anthropic’s powerful model Mythos in some cybersecurity scenarios, a development poised to reset the global tech race and pressure the White House in its overhaul of U.S. AI policy.
My lord, I hate the rules of bounded distrust, because the headline is allowed to lie (as it does here) and this first paragraph does not actually break journalistic norms.
Why? Technically speaking, yes, ‘in some cybersecurity scenarios’ Z.ai’s GLM-5.2 can match the performance of Mythos. Those scenarios could be called ‘the easy ones.’
Security researchers said that a new AI model, released this month by China’s Zhipu AI, also known as Z.ai, can match the latest U.S. models when it comes to finding security bugs, although it still lags behind Anthropic’s and OpenAI’s products in other tasks.
More precisely, this is the same finding that was harped on previously about how various models, not only GPT-5.5 and Opus 4.8 but also some open ones, could always identify any given security issue in code, provided you give them sufficient resources and point them at the particular correct subsection of code.
I would challenge the idea that GLM-5.2 can match Mythos in security bug identification in practice, but again if we sufficiently narrowly define it, then most of the things that Mythos can find GLM-5.2 can, if pointed correctly, also find.
Onward to more skirting of the line. It’s almost an art form here.
Robert McMillan, Raffaele Huang and Amrith Ramkumar (WSJ): Overall, the capability gap between top U.S. models and those built by Chinese companies has narrowed significantly, and use of Chinese AI systems has surged as businesses seek to rein in runaway costs. A host of companies, including Microsoft, are weighing how they can offer Chinese models on their platforms, a development that is set to alter the balance of power among tech companies.
Use of AI systems of all types has surged, so Chinese model usage has as well. And yes, Microsoft considered using a Chinese model for Copilot, because Copilot is the enshittified product they offer and they want to save money. And ‘set to alter the balance of power’ does not have a magnitude attached, so it technically is true.
Then there’s the claim that the ‘capability gap has narrowed significantly,’ which is true in the sense that GLM-5.2 is considerably better than previous Chinese models, which means the gap shrunk a bunch the day of its release. The observed gap is going to continuously move both up and down, and there will be some troughs along the way in terms of nowcast gaps.
The impression here and later, that the gap is predictably shrinking over time, is simply false. Before GLM-5.2, I would say the gap had clearly gotten larger, both in terms of number of cycles and also absolute clock time, since DeepSeek’s R1 and the DeepSeek moment, which itself was not especially close.
Robert McMillan, Raffaele Huang and Amrith Ramkumar (WSJ): “China is making sure that the gap becomes smaller and smaller over time,” said Lior Div, chief executive officer of the cybersecurity company 7AI.
You’re always allowed to quote someone else, but Lior Div is incorrect here.
Robert McMillan, Raffaele Huang and Amrith Ramkumar (WSJ): On Wednesday, the Chinese cybersecurity company 360 Security Technology released a new bug-finding tool called Tulongfeng. The company said it was comparable to Mythos in finding bugs. Those capabilities have alarmed many national-security officials and CEOs.
Again, the company presumably did say this. But why would you believe them? This kind of false or heavily misleading claim is made all the time.
One Helpful Note
I do want to note one helpful thing that was said:
Robert McMillan, Raffaele Huang and Amrith Ramkumar (WSJ): “Banning Fable while selling chips China needs to develop its own version is a gift to China,” said Saif Khan, a distinguished technology fellow at the Institute for Progress think tank who worked on export restrictions in the Biden administration. The U.S. needs to maximize the use of Mythos and comparable models to harden its cyber defenses while it can, he added.
Okay, yeah, that paragraph is accurate and important, so at least there’s that.
The Overall Impression Is Extremely Wrong
But the overall impression of the article, even excluding the headline, is that China is steadily catching up if it has not already done so, and even that it has matched Mythos where it counts and in the capabilities that caused the White House to delay and restrict its release, and that Chinese open models are steadily taking over AI usage. All of this is completely wrong.
All Of This Has Happened Before And Will Happen Again
This was not the first rather importantly false headline in the WSJ about AI.
See ‘Anthropic Urges Global Pause in AI Development, Flags ‘Self-Improvement Risk,’ from June 4, 2026. Anthropic came very short of any such suggestion. In that past case, the body of the article was fine, it was only the headline that was false, but that is one hell of a false impression to put onto readers.
We also have had to suffer quite a lot of ‘Chinese models have caught up’ headlines, from various sources. Some day one of them might be true. Today is not that day.
The Wall Street Journal printed an outright false headline and heavily misleading story claiming this, which of course was uncritically amplified by the usual suspects.
I post this now on its own so that we have a place to link to, to explain the situation.
Headline News
That. Did. Not. Happen.
The post even claims, explicitly, that Claude Opus 4.8 similarly ‘matches’ Claude Mythos, a claim which is even more obviously false.
Shame upon the Wall Street Journal. I fear Gell-Mann Amnesia. If they can get something as important as this so completely wrong, what about everything else?
I am skipping over the parts that involve accurate reporting, or minor quibbles.
It seems important to focus on clearly debunking the central false claims.
Alas, the mistakes made here very much rhyme with mistakes being made throughout all this by the White House, and that get latched onto by certain bad actors, who have played a large part in leaving us unprepared for the Mythos Moment.
For a full understanding of GLM-5.2, which is indeed an impressive open model, here is my full coverage of that release, placing it in proper context.
It is important to understand what makes Mythos special. This is not it.
What Makes Mythos Special
What makes mythos special is not that only the chosen one can identify any given vulnerability in code.
What makes Mythos special is that it can identify vulnerabilities autonomously, at scale, without being pointed at them, and can then autonomously string together a variety of seemingly unrelated vulnerabilities into full working exploits.
This is the thing that GPT-5.6 Sol cannot do, and that Opus 4.8 and GPT-5.5 cannot do, and that GLM-5.2 also cannot do, at anything like the same level of difficulty. It is also a thing that Fable cannot do, not with any known prompting strategy, whether or not you call this a ‘jailbreak.’
Mythos cannot do anything that you could not eventually do with a less capable model, in the same sense that an infinite number of monkeys can write Shakespeare.
The public would net benefit if both Fable and Sol were generally available in their current forms as soon as possible, including on the question of cybersecurity. Whereas releasing Mythos to the public at this time would in expectation be a serious error.
Going Over The Detailed Claims
My lord, I hate the rules of bounded distrust, because the headline is allowed to lie (as it does here) and this first paragraph does not actually break journalistic norms.
Why? Technically speaking, yes, ‘in some cybersecurity scenarios’ Z.ai’s GLM-5.2 can match the performance of Mythos. Those scenarios could be called ‘the easy ones.’
More precisely, this is the same finding that was harped on previously about how various models, not only GPT-5.5 and Opus 4.8 but also some open ones, could always identify any given security issue in code, provided you give them sufficient resources and point them at the particular correct subsection of code.
I would challenge the idea that GLM-5.2 can match Mythos in security bug identification in practice, but again if we sufficiently narrowly define it, then most of the things that Mythos can find GLM-5.2 can, if pointed correctly, also find.
Onward to more skirting of the line. It’s almost an art form here.
Use of AI systems of all types has surged, so Chinese model usage has as well. And yes, Microsoft considered using a Chinese model for Copilot, because Copilot is the enshittified product they offer and they want to save money. And ‘set to alter the balance of power’ does not have a magnitude attached, so it technically is true.
Then there’s the claim that the ‘capability gap has narrowed significantly,’ which is true in the sense that GLM-5.2 is considerably better than previous Chinese models, which means the gap shrunk a bunch the day of its release. The observed gap is going to continuously move both up and down, and there will be some troughs along the way in terms of nowcast gaps.
The impression here and later, that the gap is predictably shrinking over time, is simply false. Before GLM-5.2, I would say the gap had clearly gotten larger, both in terms of number of cycles and also absolute clock time, since DeepSeek’s R1 and the DeepSeek moment, which itself was not especially close.
You’re always allowed to quote someone else, but Lior Div is incorrect here.
Again, the company presumably did say this. But why would you believe them? This kind of false or heavily misleading claim is made all the time.
One Helpful Note
I do want to note one helpful thing that was said:
Okay, yeah, that paragraph is accurate and important, so at least there’s that.
The Overall Impression Is Extremely Wrong
But the overall impression of the article, even excluding the headline, is that China is steadily catching up if it has not already done so, and even that it has matched Mythos where it counts and in the capabilities that caused the White House to delay and restrict its release, and that Chinese open models are steadily taking over AI usage. All of this is completely wrong.
There is a reason that Tim Hua was among those having very unkind words here.
All Of This Has Happened Before And Will Happen Again
This was not the first rather importantly false headline in the WSJ about AI.
See ‘Anthropic Urges Global Pause in AI Development, Flags ‘Self-Improvement Risk,’ from June 4, 2026. Anthropic came very short of any such suggestion. In that past case, the body of the article was fine, it was only the headline that was false, but that is one hell of a false impression to put onto readers.
We also have had to suffer quite a lot of ‘Chinese models have caught up’ headlines, from various sources. Some day one of them might be true. Today is not that day.
Thank you for your attention to this matter.