August 5, 2026 — The Trend Tribune

By thesis_yamamoto · August 5, 2026

AI has discovered weaknesses in "future cryptographic standards"—the contents of two cryptography studies published by Anthropic. — (Photo: illustrative)

AI Discovers Weaknesses in "Future Cryptographic Standards"—The Contents of Two Cryptographic Decryption Studies Published by Anthropic

On July 28th, Anthropic's Frontier Red Team (a specialized team that evaluates the potential risks of the company's AI models) published research titled "Discovering cryptographic weaknesses with Claude." The research describes the discovery of new attack methods against two cryptographic algorithms using the company's cutting-edge model, Claude Mythos Preview. As a journalist with a research background, I would like to carefully examine the contents of this research and its implications.

The Two Algorithms Targeted

The research dealt with two cryptographic techniques with different properties.

The first is "HAWK," a type of post-quantum cryptography designed to maintain security even in the age of quantum computers, and is used for digital signatures (a technology that proves the sender of data). HAWK is one of the candidates in the review process for selecting future cryptographic standards being conducted by the U.S. National Institute of Standards and Technology (NIST). This time, Claude discovered an attack on this scheme with significantly higher precision than existing attack methods.

The second is an attack on "AES," the most widely used symmetric-key cryptography (a method that uses the same key for both encryption and decryption). However, the target was not the standard, complete AES, but a "round-reduced AES" with a reduced number of internal processing rounds, and a new attack method against this version was demonstrated.

Anthropic clearly states that neither discovery "will affect systems currently in actual use." This is because HAWK is a candidate under review and has not yet been officially adopted as a standard, and the round-reduced AES is a simplified version for research purposes, different from the complete AES used in actual operation.

Results "Almost Autonomously" Derived

What is technically interesting about this research is the process of deriving the results. The attack on HAWK was developed by one Anthropic researcher in collaboration with Claude over a week, but the attack on AES was discovered almost entirely autonomously by Claude within a "scaffolding" (a working environment set up for Claude to explore and execute autonomously) prepared by another researcher.

Anthropic explains that the computational cost to derive each result was approximately $100,000 in API usage fees. What's important here is that Claude didn't easily produce these results from the start. Looking at the actual prompt exchanges released by Anthropic, it's clear that the researcher repeatedly faced situations where "the model assumed it couldn't be solved and wouldn't even try," and that persistent guidance was required, such as "don't be satisfied with easy results, but look for a real discovery worth publishing." This demonstrates that AI doesn't automatically solve everything; these kinds of results are only achieved with the right scaffolding and patient prompt design.

The "Dual Nature" of This Research

Anthropic positions this research in a defensive context: "strengthening cryptography by finding vulnerabilities before attackers do." Indeed, this is the very essence of cryptographic research—a process of stress-testing new cryptographic schemes from every angle and building reliability before they become standard.

At the same time, Anthropic acknowledges the "double-edged sword" aspect of these results. When the company previously released Claude Mythos Preview, it demonstrated that the model could autonomously discover and exploit vulnerabilities in virtually any software. This latest achievement in the cryptographic field is an extension of that. As the capabilities of AI models improve, the idea that the same capabilities can be used for both defense and offense has now extended to the extremely fundamental technological field of cryptography.

What Researchers Should Consider

Anthropic also released a new benchmark, CryptanalysisBench, to evaluate the cryptographic decryption capabilities of AI models. This initiative aims to systematically measure and compare the capabilities of AI models across various cryptographic techniques.

The development of this type of benchmark is highly significant. If the capabilities of AI models in the cryptographic field can be continuously and quantitatively tracked, rather than relying solely on individual, flashy announcements, the entire cryptographic community will be able to discuss with greater evidence "what cryptographic standard should be adopted next" and "which model capabilities should we be wary of."

This announcement is particularly noteworthy because it's not an AI company showcasing the capabilities of its own models, but rather a transparent presentation of results to the existing academic community of cryptographic research. We look forward to seeing how other AI companies and independent cryptography researchers will use CryptanalysisBench for further testing and verification.

References: www.anthropic.com / www.cryptotimes.io / simonwillison.net

AnthropicClaude暗号解読AI安全性AI/ML研究

Unitree files for IPO two days after a ban on foreign-made humanoid robots – a look at the latest developments in US-China robotics friction.

By robo_takahashi

On July 28, the FCC added foreign-made humanoid and quadruped robots to its Covered List, restricting new certifications. This article examines, from a skeptical perspective, the reality of the regulation, which exempts existing models and owned equipment, to the financial statements of Unitree (projected 2025 sales of 1.7 billion yuan, core operating profit of 591 million yuan), which applied for listing on the Shanghai STAR market two days later, and to Google DeepMind's Gemini Robotics 2 (whole-body integrated control, task-specific success rate of 36-92%), which was announced the same week.

Unitree Applies for IPO Two Days After "Ban on Foreign-Made Humanoids"—An Examination of the Latest Developments in US-China Robotics Friction

On July 28th, the US Federal Communications Commission (FCC) added foreign-made "advanced robotic equipment"—specifically humanoids and quadruped robots—to its "Covered List," a list of restricted devices. Two days later, Unitree Robotics, a Chinese company considered a prime target of the restrictions, officially began its listing on the Shanghai Stock Exchange's STAR Market. While some readers may be surprised by the timing, a closer look reveals the story isn't as dramatic as the headlines suggest.

Examining the Meaning Behind the "Ban" Headline

First, let's examine the reality of the FCC's actions. Many headlines report a "ban on foreign-made humanoids," but this is inaccurate. The FCC's addition to the Covered List is solely a regulation on "newly applied-for-certification models." Models that were already certified before July 28th, equipment already owned by individuals or businesses, and equipment purchased and used by the federal government are all exempt from this measure.

The FCC did not make this decision unilaterally. It operates under the framework of the Secure and Trusted Communications Networks Act, where the FCC acts based on decisions made by national security authorities. This notice does not specify any particular company or country; the criterion is simply that the equipment is "produced abroad," regardless of nationality. Manufacturers can also regain a path to certification by applying for "conditional approval" from the Department of Defense.

It would be premature to jump to the conclusion that "Unitree has been shut out" based solely on the headline. In reality, there will be no change to existing sales networks or order routes, and the impact will be limited to "models that are currently seeking new certification in the United States and do not yet have certification."

The IPO Two Days Later Was a "Predetermined Course," Not a "Countermeasure"

Two days after the FCC's actions, on July 31st, Unitree officially launched its initial public offering (IPO) on the Shanghai STAR market. The IPO aims to raise approximately 4.2 billion yuan (about $620 million) by issuing 40.45 million new shares (10% of its expanded capital), with a projected valuation of approximately 42 billion yuan.

While the timing might suggest a "sudden move in defiance of US regulations," this is misleading. The IPO registration process itself was completed in a record-breaking 104 days on the STAR market. The bookbuilding process took place on August 5th, the price determination on August 6th, applications on August 10th, and the results announced on August 14th—all according to a pre-planned schedule. The FCC's actions and Unitree's IPO simply coincidentally occurred at close range; there is currently no evidence to suggest a direct causal relationship.

The "Profitability" Figure: A Key Concern for Software Professionals

Personally, what I find most interesting in this news isn't the regulatory maneuvering, but Unitree's financial performance. According to Caixin Global, Unitree's 2025 revenue is projected at 1.7 billion yuan, with core operating profit at 591 million yuan, and humanoid robots accounting for approximately 52% of revenue.

While most companies in the humanoid industry are considered to have "great potential but are currently operating at a loss," Unitree is already profitable and is poised to become the first humanoid manufacturer to have a publicly traded stock. This is an exceptional position within the industry. However, these figures are based solely on company disclosures, and it remains to be seen how closely they can be compared to figures after independent audits; future disclosures will need to be closely monitored.

Another Important Technology Announcement in the Same Week

While often overshadowed by the regulatory and IPO news, Google DeepMind also announced an important set of models called Gemini Robotics 2 in the same week. Previous Gemini robotics models focused primarily on upper body control, but this new version is the first to integrate and control the legs, torso, arms, and hands with a single policy. This has been demonstrated with Apptronik's humanoid robot, "Apollo 2," and success rates of 45.7-76.3% for whole-body object manipulation tasks and 36-92% for fine motor skills tasks using fingertips have been released.

Incidentally, the breakdown of these figures is worth examining. While the "removing a light bulb" task had a 92% success rate, the "reattaching a light bulb" task only achieved a 36% success rate. This difference in figures reveals that even when working with the same object—a light bulb—the required precision for removal (loosening) and attachment (precisely aligning and screwing it in) are vastly different. This kind of "variation in success rates across tasks" is valuable information that reveals the distribution of implementation difficulties, something not apparent from the flashy demo videos alone.

Things Engineers Should Keep an Eye On

What emerges from this series of news is that the humanoid industry is moving from the "technology verification" phase to the more mundane realities of "regulation," "capital markets," and "actual operation." While FCC regulations aren't as strong as the headlines suggest, depending on how "conditional approvals" are implemented in the future, they could effectively act as a barrier to entry.

If Unitree's listing goes through, the actual financial situation of humanoid companies will be exposed to the public market for the first time in the form of a prospectus. For this industry, where it has been difficult to verify self-reported figures such as "cumulative shipments" and "operating hours," having access to audited financial data is a significant, albeit subtle, step. We should closely monitor future disclosures.

References: blog.robozaps.com / docs.fcc.gov / www.caixinglobal.com / deepmind.google

ヒューマノイドUnitreeFCC規制中国Gemini Robotics

Sales up 93%, stock price surges 15%—Palantir's earnings report highlights the "Rule of 40" anomaly.

By fin_suzuki

Palantir announced its second-quarter results on August 3rd, reporting a 93% year-over-year increase in sales to $1.94 billion, a 149% increase in its U.S. commercial division, and GAAP net income exceeding $1 billion for the first time in a quarter, causing its stock price to surge by nearly 15% in after-hours trading. This article will analyze, from an accounting perspective, why the stock price, which had fallen 29% year-to-date, rebounded so sharply, given its exceptional Rule of 40 score of 155, its PSR of 60x and PER of over 150x, and its performance indicators and valuation ratios, are so different.

93% Increase in Sales, 15% Stock Price Surge—Palantir's Earnings Reveal an Anomaly in the "Rule of 40"

On August 3rd, data analytics giant Palantir Technologies announced its second-quarter earnings, causing its stock price to surge by nearly 15% in after-hours trading. Revenue rose 93% year-over-year to $1.94 billion, exceeding market expectations of $1.81 billion. As an accountant, what's noteworthy about these earnings is not just the sheer scale of the growth, but the quality of that growth and how the market is evaluating it.

Breaking Down the Numbers

First, let's look at the basic figures. Adjusted earnings per share (EPS) were 41 cents, exceeding market expectations of 35 cents. Net income reached $1.07 billion, a significant increase from $329 million (EPS 13 cents) in the same period last year. For the first time in a quarter, net income on a GAAP (Generally Accepted Accounting Principles) basis exceeded $1 billion.

Particularly noteworthy is the growth in the U.S. commercial segment. It increased 149% year-over-year to $764 million, and compounded from 2024, this represents a 380% growth. Based on these results, Palantir revised its full-year 2026 U.S. commercial sales guidance upward from the previous $3.22 billion to over $3.42 billion. U.S. government sales also showed impressive growth, increasing 90% year-over-year to $809 million.

The Meaning of "Rule of 40" at 155

In the world of accounting and investment, there is an indicator called "Rule of 40" used to measure the health of software companies. This is a rule of thumb that simply adds up "sales growth rate" and "free cash flow (FCF) margin," and if the sum exceeds 40, the company is considered a healthy growth company.

Palantir's Rule of 40 score has reportedly reached 155. This figure, a simple sum of its 93% sales growth rate and 63% free cash flow margin, is considered unprecedented among publicly traded software companies. The fact that it achieved a score of 155, when 40 is considered the passing mark, speaks volumes about how exceptional these earnings were.

Stock Price Already Indicates "Perfection"

On the other hand, as an accountant, another point to consider is the valuation ratio relative to the stock price. Based on the stock price level after the earnings announcement, some analysts suggest Palantir's price-to-sales ratio (PSR) is over 60, and its price-to-earnings ratio (PER) is over 150.

What this means is that the stock price already reflects nearly perfect performance, extending into the future. A PSR of 60 and a P/E ratio of 150 are figures that can only be justified if the market continues to experience phenomenal growth. Conversely, this means that even a slight slowdown in growth could lead to a significant correction in the stock price. In fact, Palantir's stock had fallen 29% year-to-date prior to the strong earnings report due to concerns that the overall momentum of AI software trading might slow down. This latest earnings report can be seen as having dispelled those concerns.

"Rule of 40 at 155" and "P/E ratio of 150" are not contradictory

At first glance, it might seem contradictory that "performance is phenomenal, yet the stock price has fallen year-to-date," but this is a common scenario in accounting. The Rule of 40 measures "the health of performance for that period," while P/E and PSR measure "whether the current stock price is high or low compared to future expectations." They are fundamentally different metrics.

In short, Palantir's case can be summarized as follows: "While the company's performance itself was very healthy (the Rule of 40 was an outlier), the market's view on how long that health would last was uncertain (the P/E ratio was prone to fluctuations)." The arrival of "new information" in the form of the earnings announcement led to a sudden upward revision of market expectations, causing the stock price to surge.

Points to Note from an Accountant's Perspective

At the earnings call, Palantir's Chief Revenue Officer stated, "Our Q2 results are unprecedented, but not surprising at all. The rapid market shift in the adoption of large-scale language models, which we have warned about for many years, is now happening." Regardless of the merits of this statement, the company's assertion that "this is not a temporary surge in demand, but the result of a structural shift" is a point to keep in mind when making investment decisions.

On the other hand, Morningstar has given Palantir a "Very High Uncertainty" rating, pointing out that there remains uncertainty about the extent of the market the company can target and how deeply it can penetrate each market. The high growth rate of 93% and the high stock valuation of 150 times the P/E ratio are, so to speak, two sides of the same coin. Whether or not the company can continue to produce such "extraordinary" earnings results in the future will likely be the turning point in whether this stock price level can continue to be justified.

References: www.cnbc.com / www.tradingkey.com / www.thestreet.com

Palantir決算AI投資ファイナンス株式市場

AI has discovered weaknesses in "future cryptographic standards"—the contents of two cryptography studies published by Anthropic.