claude

At the end of September and the beginning of October, at least seven financial institutions in South Korea experienced data breaches, including Shinhan Bank, Kookmin Bank, and Hana Bank. Approximately 25,000 customers of Shinhan Bank and about 40,000 customers of Yegaram Savings Bank were affected, with leaked information including names, phone numbers, annual income, and loan limits; no theft of funds has been discovered so far.The attacks primarily targeted loan inquiry services used by loan intermediaries and peripheral banking systems such as employee mobile office systems. Multiple attacks involved overlapping IP addresses, leading South Korean regulatory authorities to suspect they were carried out by the same attacker. On October 6, South Korean President Yoon Suk-yeol stated that these attacks may have utilized AI.On October 7, CrowdStrike reported that the attackers controlled the attack using a server located in Hong Kong, with another running the open-source AI penetration testing system ARTEX. Investigators obtained the ARTEX configuration files, Claude Code chat records, and AI memory files. The configuration indicated that the attackers primarily used DeepSeek v4.1-flash to drive ARTEX, also utilized Claude Code, and had previously called GLM-5.3 and Grok 4.6 in other sessions. The chat records showed that the attackers inquired about channels for selling the leaked data in South Korea and related Telegram trading groups, and instructed the AI to draft a resume for a security researcher that included information such as 26 years old, Maoming, Guangdong, South China University of Technology, and contact details, but the birth date did not match the age. CrowdStrike assessed that the attackers might be using Chinese and aimed for profit, but their true identity has not yet been confirmed.

Anthropic released Claude Haiku 5.5 on Wednesday, claiming it to be the company's cheapest, fastest, and most capable small model to date. This model is designed for high-volume tasks such as document summarization and database queries, as well as speed-sensitive scenarios like real-time customer service and browser automation. When developers integrate it into their own products, prompts of up to 100,000 tokens are charged at $0.1 per million input tokens and $0.5 per million output tokens, consistent with the pricing of OpenAI's competing small model GPT-6 Luna released on September 22.The corresponding prices for Haiku 4.5 are $1 and $5, with the new prices being 90% lower, and prompts exceeding 100,000 tokens receiving a 50% discount. Since about 90% of requests for the old model fall below this threshold, and Haiku 5.5 has a slightly higher token count, Anthropic estimates an average savings of about 75%. Benchmark tests show that Haiku 5.5 achieved a success rate of 72.4% on OSWorld 2.1, higher than GPT Luna's 48.9%; on Terminal-Bench 4, it scored 39.2%, surpassing Luna's 16.4% and Haiku 4.5's 0%, but lower than Sonnet 5.5's 70.6%.This is the first Haiku model with adjustable effort settings, allowing users to balance cost and answer quality.

According to Decrypt, Carli Michelle Heller, a woman from Bonita Springs, Florida, was arrested for using Claude as a diary. Investigators stated that she wrote about "shooting" the Lee County Sheriff's Office on September 26, and the next day claimed she had acquired a new gun. Anthropic's safety system flagged these conversations as serious, and the platform handed them over to a human review team, which reported to law enforcement. Officers then identified Heller, went to her residence, and detained her. She was charged with making written threats of violence, and the case is scheduled to go to court in November.Anthropic's consumer terms explicitly reserve the right to "decide on their own" to report user inputs, outputs, or behaviors to law enforcement. Its privacy policy, effective September 10, 2026, also allows the company to share data with police when it reasonably believes that disclosure is necessary to prevent serious harm. Even if users choose not to use their data for model training, Anthropic can still use their data within three months, rather than five years; conversations flagged for review will also be used to improve harmful content detection capabilities.This is not the first case. The San Francisco Standard reported in September that Anthropic had referred a user to the police, who allegedly wrote that they purchased an AR-15 rifle and had the company's CEO Dario Amodei in their sights, claiming it was just a joke. At the time of the report, the user had not been arrested or charged, and an Anthropic spokesperson stated that this was "our protective process working as intended."

According to official news, WasabiCard has launched various AI subscription cards, catering to the global AI tool subscription and payment needs of enterprises and teams, supporting subscription payment scenarios for mainstream AI applications such as OpenAI and Claude. Through card-based payment solutions, companies can handle the procurement and expense management of AI services for different teams, projects, or business lines more flexibly.In addition to AI subscription payments, WasabiCard also supports bulk card issuance for enterprises, fund budget management, and multi-currency account setup, as well as various scenarios such as advertising, business travel, payroll, and e-commerce subscriptions. It helps companies configure payment permissions and budget limits by department, employee, or specific purpose, enhancing the management efficiency of AI-related expenditures and the transparency of fund usage.As AI tools gradually enter daily workflows in research and development, content production, customer service, and operations, cross-border subscription payments and expense management are becoming important infrastructure for global operations of enterprises. It is important to note that currently, WasabiCard-related products are not open for application to enterprises and individual users in mainland China.

On September 29, 2026, the Anthropic Frontier Red Team published a research article authored by Andrew Fasano, Marius Fleischer, Cole McFaul, Robert Xiao, and Tripp Gallagher. The article states that about five months ago, Anthropic released the Claude Mythos Preview through Project Glasswing in a limited manner, as it can autonomously construct end-to-end exploit capabilities; trusted defenders discovered over 10,000 vulnerabilities in critical software based on this, and similar capabilities have now spread to other models.The article mentions that the GLM-5.3 developed by ZhiPu AI (known overseas as Z.ai) is similar to the Claude Mythos Preview, possessing strong autonomous end-to-end exploit development capabilities, but it is released with open weights and lacks meaningful abuse restrictions. In simulated tests, the success rate of bypassing GLM-5.3 defenses using simple methods was about 64% to 100%, while the same methods failed to make the protected Claude model perform harmful tasks; these capabilities can also be used by defenders to strengthen systems.On September 17, 2026, the National Institute of Standards and Technology's Artificial Intelligence Standards and Innovation Center assessed that GLM-5.3 is currently the most powerful open-weight model in terms of network capabilities, lagging behind the U.S. frontier by about four months in its network benchmarks, and Anthropic stated that the capability conclusions are generally consistent.

On September 29, B.AI announced that Anthropic's latest Sonnet model, Claude Sonnet 5.5, has officially launched on the platform. As a faster and lower-cost supplement to Claude Opus 5.5, Claude Sonnet 5.5 boasts an output speed increase of over 30%, with a maximum reduction in single-task costs of about 30%, aimed at clearly defined software engineering, everyday agent workflows, and professional document creation scenarios.This model supports image and PDF comprehension, can generate reports, presentations, spreadsheets, and UIs, and offers five adjustable reasoning intensities with a 1 million token context window, with no additional charges for long contexts. Currently, both the B.AI API and Web Chat have been synchronized and opened, allowing developers and users to directly access this new generation Sonnet model that balances speed, cost, and long context capabilities.

Anthropic released Claude Sonnet 5.5 this Monday, as an upgraded version of Sonnet 5 launched in June this year. The official statement claims that this mid-range model operates over 30% faster than the previous generation, excelling in clearly defined daily tasks, fixing defects, and creating documents, slides, and spreadsheets, while also possessing sharp design capabilities. Its pricing remains at $2 per million input tokens and $10 per million output tokens, which is half of Opus 5.5, and the token consumption per task is nearly one-third less than Sonnet 5.In the Terminal-Bench 4 test, Sonnet 5.5 scored 70.6%, higher than Opus 5.5's 66.4%, and Sonnet 5's score was only 10.3%. Results from the independent testing organization Artificial Analysis also show its lead, recording 63.6% and 59.6%, while OpenAI's GPT-6 Astra scored 59.1%. In the GDPval-AA assessment of real job performance across 44 occupational categories, Sonnet 5.5 received a score of 1844, which is roughly on par with Opus 5.5's score of 1846, while GPT-6 Sol scored 1487.

On Wednesday local time, Anthropic announced that its model Claude identified a previously undescribed enzyme system in vast DNA sequence data, with a repetitive sequence arrangement similar to CRISPR. This is the first achievement of its newly established biological laboratory, which was formed this spring and is located in the San Francisco Bay Area. Following the announcement, shares of gene editing companies fell on Wednesday, with CRISPR Therapeutics down 5.54%, Beam Therapeutics down 6.02%, and Prime Medicine down 11.58%.The related task was to search for noteworthy reverse transcriptase systems from large DNA databases. During approximately 21 hours of searching, about 950 Claude agents used around 210 million tokens, with one agent noticing a repetitive DNA sequence arranged near a certain reverse transcriptase gene. The team named this combination the array-associated reverse transcriptase system, primarily found in bacteriophages, which includes reverse transcriptase genes, adjacent accompanying genes, and a set of repetitive DNA sequences. The reverse transcriptase itself has been studied before, and the new finding is that it forms a system together with surrounding non-coding DNA sequences and functionally unknown accompanying proteins.Anthropic stated that researchers still do not know the main function of this system and have not proven that it can perform targeted modifications to DNA like mature gene editing tools. Preliminary experiments found that these repetitive sequences can produce various short RNAs, but this does not equate to having programmable gene editing capabilities. The preprint paper disclosed that to verify whether the discovery could be repeated, the team ran the same search 10 more times;

According to The Information, citing informed sources, China's National Internet Information Office has launched an investigation into AI companies DeepSeek and Moonshot AI, triggered by Anthropic's allegations that the two companies secretly routed sensitive user data to their servers. Reports indicate that regulators visited the offices of both companies, interviewing executives and employees, focusing on whether sensitive data related to law enforcement, military, and state-owned enterprises has flowed into U.S. servers.The trigger for this investigation was Anthropic's fourth threat intelligence report released on September 10. This 154-page document accuses seven Chinese labs—Alibaba, Moonshot AI, DeepSeek, Z.ai, MiniMax, SenseTime, and Xiaomi—of engaging in what it calls "illegal distillation," which involves using the outputs of large models to train smaller models. Anthropic states that distillation itself is a legitimate practice, but it opposes its implementation through fraudulent accounts. The National Internet Information Office initially summoned all seven companies named in the report, but later narrowed the investigation to DeepSeek and Moonshot AI. Anthropic claims that Moonshot AI routed over 23 million interactions to Claude through 5,380 fraudulent accounts, while DeepSeek generated over 12.1 million interactions within a 14-day window in July.The timing of the investigation is quite delicate for both companies.

On September 23, the B.AI platform announced that the first model of Anthropic's newly released Claude 5.5 series, Claude Opus 5.5, is officially online, with API and Web Chat both open for simultaneous experience. This model is designed for long-cycle Agent programming and complex knowledge work, achieving Claude Fable 5.1 levels in most tasks, while reducing operating costs by about 40% compared to Opus 5.The model supports 1M Token ultra-long context and up to 128K output, demonstrating leading capabilities in Computer Use and multi-step automation scenarios. From today, developers and users can directly access Claude Opus 5.5 through the B.AI platform. You are welcome to log in to chat.b.ai/chat to experience it immediately.

On Tuesday, Anthropic released Claude Opus 5.5, the first model in its Claude 5.5 series. Anthropic stated that the new model's performance on most tasks has reached the level of its most expensive flagship model, Claude Fable 5.1, but the operating costs are 20% lower than the previous generation Opus 5 and 60% lower than Fable 5.1. In terms of pricing, the input and output token prices for Opus 5.5 are $4 per million and $20 per million, respectively, while the cache read price has dropped to $0.2 per million.In benchmark testing, Opus 5.5 scored 66.4% on Terminal-Bench 4, higher than Fable 5.1's 55.8% and GPT-6 Astra's 57.9%; FrontierCode scored 54.4%, higher than Fable 5.1's 50.3%; GDPval-AA v2.1 scored 1846, higher than Fable 5.1's 1735. However, GPT-6 Astra still leads Opus 5.5 in the AutomationBench and Terminal-Bench-Science 0.1 tests.This is the first model released after Anthropic CEO Dario Amodei called on the industry to slow down the pace of AI capability enhancement, and it continues to use the same cybersecurity and biosafety measures as Fable 5.1.

PPP market prediction tool monitoring shows that the probability of "Polymarket's next Claude Opus model being released on September 22" has risen to 77%, up 50% in 24 hours. Only the new model officially named "Opus" by Anthropic counts, such as Claude Opus 5.1, Opus 5.5, Opus 6, etc. Other names like Sonnet, Haiku, Fable, Mythos do not count unless officially named Opus by Anthropic.The release time is based on the calendar date in Eastern Time, USA. It must be genuinely open to the public to trigger settlement. Public beta and open rolling waitlists count; closed beta, invitation-only, or private access do not count. If the model name, placeholder, or labeling error appears only on Anthropic's official website, but the model is not actually available for public use, it does not count as released. The final settlement will primarily rely on official information from Anthropic, verified with credible media reports.

Elon Musk retweeted a post stating that GPT-6 Astra pushed a simulated character off a cliff in multiple simulation tests, while Grok, Gemini, and Claude did not do so.

PPP prediction market tool monitoring shows that the probability of Polymarket "Anthropic's next generation Claude Opus will be released before September 27" has risen to 90%, up 13% in 24 hours; the probability of it being released before September 30 has risen to 94%, up 7% in 24 hours.According to the settlement rules, if Anthropic's next Claude Opus model is made publicly available before the specified date (Eastern Time ET), the prediction market will settle as "Yes"; otherwise, it will settle as "No". Claude Opus refers specifically to the model explicitly named "Opus" by Anthropic; models released under other names, such as Sonnet, Haiku, Fable, or Mythos, do not qualify. Additionally, qualifying models must have been officially launched and made publicly accessible, including open testing or public rolling waitlist registration. Closed testing or any form of private access does not qualify for settlement.

Anthropic stated that its chatbot Claude leads 26% of the company's artificial intelligence research and development work, a significant increase from nearly zero at the beginning of this year. The report released by the company on Thursday noted that this is one of the clearest signs to date that artificial intelligence can accelerate the development of future models.Anthropic mentioned that Claude also collaborates with employees to complete about 90% of the work. The company pointed out that more than a quarter of the artificial intelligence research and development is driven by Claude.

According to The Wall Street Journal, an independent security research team used Anthropic's Claude software to breach a ChatGPT account belonging to an OpenAI employee, allowing them to access and modify the internal code system associated with that employee.The researchers first used Claude to analyze a vulnerability in Discourse, which is used by the OpenAI developer community, and generated executable attack code. They then obtained an authentication token and, due to a permissions configuration issue, accessed a ChatGPT account belonging to an OpenAI employee, while also gaining limited read and modification suggestion permissions for some private GitHub repositories.

Mustafa Suleyman, the head of artificial intelligence at Microsoft, warned that injecting human-like characteristics into tools like Anthropic's Claude increases the risk of such systems going out of control.In an article published on Wednesday, Suleyman criticized certain wording in the Claude guidelines. The charter of Claude expresses a vague attitude about whether the assistant is an ethical entity and assumes that the software may have some version of functional emotions or feelings.

Anthropic announced that starting today, Claude Cowork and chat will merge into the same Claude. Users can ask brief questions or deliver deadline tasks, and Claude can continue processing even after the computer is turned off. This capability will be rolled out to Pro and Max plans in the coming weeks, followed by more plans.Claude Docs and Claude Slides will launch simultaneously, and Claude Design is now available for use in conversations. Users can co-write documents or draft slides with Claude, supporting direct editing, presenting in Claude, and downloading as PowerPoint or PDF. The three features are currently in testing for paid plans, and enterprise administrators can decide when to enable them; the standalone Claude Design can still be used as usual.Anthropic stated that previously, Cowork and Design belonged to different entrances, making task attribution and context connection inconvenient. Claude can now determine the capabilities required for tasks and call upon existing context, skills, and connectors in any conversation. Chat users do not need to change their operations; Cowork users can retain chats, projects, products, connectors, and skills. Team and Free plans will be launched subsequently, and enterprise administrators will be informed of organizational changes at least 30 days in advance.

WeChat has started to further relax the circulation of chat records. Practical tests show that after upgrading the mobile WeChat to version 8.0.78, users can select multiple chat records and choose "Forward to other applications." In addition to Yuanbao and WorkBuddy, users can directly send them to third-party apps like ChatGPT by selecting "Choose applications on the phone," with a maximum of 100 records selected at once.WeChat will package the chats into a ZIP file, containing a TXT file of the chat records organized by time and related attachments. The desktop version of WeChat has also opened a similar entry, and community developers have created intermediary tools based on this, allowing chat records to be sent to AI like ChatGPT and Claude.