LectureRechercheRadarCadre d'investissement
Connexion / Inscription
ToutTechnologieProduitsInvestissementBourseAutres

Lecture

2026-11-10115 publications

Actualiser

UK Government Code Exposes Over 100 Vulnerabilities Found by AI Models, Sparking Open-Source Policy Debate

Sujet · 英国政府代码开源安全争议Matériel
2026-11-10 15:00 GMT+8

Security researcher Chris Nesbitt-Smith from ControlPlane used Anthropic's Claude Opus and Sonnet AI models to scan publicly available codebases across multiple UK government departments, discovering more than 100 serious vulnerabilities. These flaws put sensitive data at risk due to poor cyber hygiene rather than sophisticated hacking techniques. Nesbitt-Smith noted that he used consumer-grade models on a laptop and stated that many issues could have been caught by routine IT health checks.

The most severe case involved the Planning Inspectorate, which handles land-use matters across England. The agency failed to fully remove login credentials after accidentally committing them to a public GitHub repository. These credentials provided "god-tier" read and write access to Microsoft Entra ID, allowing access to all SharePoint sites, private OneDrive accounts, and HR records. Although the agency fixed the vulnerability within 20 minutes of being reported and verified no data breach occurred, the incident highlights neglect in maintaining technology estates.

The audit covered 349 public sector codebases across 33 organizations, with vulnerabilities found in approximately two-thirds of them. Affected entities included the Department for Education (involving school procurement services, teacher application systems, and training provider registers) and the Scottish Government's open data portal. While the UK government has defaulted to openly publishing code produced with public money for around a decade to enhance transparency and reuse, funding squeezes have led to neglected maintenance.

As AI accelerates vulnerability discovery, arguments for public sector departments to stop publishing source code have intensified. Reports indicate the NHS ordered technology teams to transition away from open source code in May amid security concerns. However, Nesbitt-Smith warned that closing repositories without fixing underlying issues is the "worst case scenario," as adversaries may have already mirrored these repositories, and closure might draw more attention to the problems.

Sources :WIRED

Recherche

Otter AI Adds Support for 27 New Languages and Automatic Language-Switching Detection

Matériel
2026-11-10 15:00 GMT+8

AI note-taking tool Otter AI announced on November 9, 2026, that it is expanding its multilingual capabilities, including a new transcription feature that can follow conversations as speakers switch between languages during the same meeting. This update marks the company's largest language expansion to date, adding support for 27 new languages. Once the rollout is complete, Otter will support 31 languages in total, making the platform more accessible to international teams.

The upcoming feature will automatically detect language changes and continue transcribing without requiring users to adjust settings. For example, a call could move from English to Korean to Portuguese, with Otter following the conversation throughout. The update aims to address a common challenge for global teams, where language differences can make meetings and collaboration more difficult.

Thirteen of the new languages are available now: Mandarin Chinese, Korean, German, Italian, Cantonese, Portuguese, Russian, Indonesian, Malay, Dutch, Turkish, Vietnamese, and Polish. The remaining 14 languages, including Swedish, Danish, Arabic, Hindi, Thai, Hebrew, Norwegian, Greek, Macedonian, Persian, Czech, Hungarian, and Romanian, are coming soon. The newly added languages run on Otter’s own speech models, which the company stated were trained using publicly available datasets and reviewed for accuracy with native-speaking volunteers.

The announcement comes as Otter continues to expand beyond transcription into other workplace AI tools, following its $320 million funding from General Catalyst. It recently introduced Work Mode, which uses meeting history and connected apps to help users prepare for upcoming calls, and Live Assist, which provides real-time coaching during meetings.

Sources :TechCrunch

Recherche

US CISA Adds Four High-Risk Vulnerabilities to Known Exploited List, Mandates Federal Patching

Sujet · CISA KEV名单漏洞Matériel
2026-11-10 14:07 GMT+8

The U.S. Cybersecurity and Infrastructure Security Agency (CISA) announced on November 9 that it has added four vulnerabilities with evidence of active exploitation to its Known Exploited Vulnerabilities (KEV) catalog. These vulnerabilities affect critical infrastructure sectors, including GPUStack for AI model deployment (CVE-2026-102895), the FreeSWITCH software-defined telecom switch (CVE-2026-116779), the OpenSSH remote connection library (CVE-2026-123158, known as "SSHole"), and IBM's Aspera Faspex file transfer platform (CVE-2026-123218).

Under CISA's directive, all federal agencies must complete patching these vulnerabilities by November 12. This tight deadline underscores the significant risk these flaws pose to government information systems if exploited. Among the four, the OpenSSH "SSHole" authentication bypass vulnerability has the broadest impact, while the other three carry a CVSS severity score of 9.8, indicating maximum-level criticality.

Sources :iThome

Recherche

Independent Analyst Richard Windsor: Jensen Huang's Damaged Credibility Undermines the Political Feasibility of Arguments for Easing China Chip Export Restrictions

Matériel
2026-11-10 13:16 GMT+8

In his latest analysis, independent technology analyst Richard Windsor presents a judgment: the controversy surrounding Nvidia CEO Jensen Huang has stripped proponents of easing AI chip exports to China of their political footing, regardless of the merit of their economic arguments. Windsor notes that it is hardly surprising that Huang has been one of the most outspoken advocates for this position—given that over 80% of the Chinese population holds a positive view of AI, the country possesses abundant electricity resources, and Nvidia arguably offers the world’s best systems for AI training and inference, a market access that is currently being compressed by its inability to sell into China.

However, Windsor argues that the crux of the matter lies not in whether the economic logic holds, but in who is presenting that logic. He writes that Huang’s fiduciary duty to Nvidia shareholders requires him to lobby the United States government to lift restrictions, which is inherently defensible. Yet, recent revelations—allegations suggesting a potential connection between Huang’s spouse and the Chinese state—have fundamentally undermined the political credibility of this lobbying path. Windsor frames this using a conditional statement:

If Ms. Huang were indeed an agent of the Chinese state, it could be easily argued that her role was to influence Jensen to push for relaxed export restrictions in Washington—which is precisely what he has done.

Windsor further analyzes the strategic logic from the Chinese side. He points out that anyone engaged in trade negotiations with China will tell you that Beijing’s primary demand is for advanced Nvidia chips, rather than the downgraded versions permitted for export by the United States. Interpreting this negotiating behavior alongside Huang’s lobbying activities leads to a specific inference: China desires advanced chips, not the second-rate alternatives the U.S. is willing to offer. The rationale behind this strategy is clear: advanced chips would significantly enhance China’s competitiveness in AI while buying time for domestic chips to catch up, ultimately aiming to remove Nvidia and other foreign vendors from its systems entirely.

Windsor concludes that the practical outcome of this affair will be a tightening of U.S. import restrictions on China, perpetuating its disadvantage in AI silicon. This implies that China will become even more dependent on importing computing services from data centers built in neighboring countries utilizing Western chips. While Windsor acknowledges this alternative is viable, he highlights a structural vulnerability: the United States could legislate to prohibit data centers using U.S. chips from processing workloads destined for use in mainland China. Although this measure has not yet been enacted, it remains a viable policy tool as long as the dependency exists. Windsor believes this dynamic explains why the United States and the West still maintain a significant advantage in AI, an edge unlikely to change in the short term.

It must be noted that Windsor’s core argument rests on unverified allegations, and he himself employs conditional phrasing. The value of his analysis lies in providing a specific mechanism explaining how personal credibility can render a policy argument—otherwise independently valid—unusable in a political environment, rather than offering independent verification of the allegations themselves. Even absent this credibility crisis, the economic logic for relaxing exports could be revived by other advocates; however, Windsor judges that under the current political climate, this path has become effectively impassable.

Sources :radiofreemobile.com

Recherche

Anthropic Finds: Declaring Code Ownership in Prompts Raises Claude Fable 5.2's Duplicate Code Merge Rate from Under 10% to About Half

Prise rapide
2026-11-10 13:34 GMT+8

Anthropic documented a finding with direct implications for coding agent behavior in the prompt engineering guide for Claude Fable 5.2: how the model handles existing code structures depends heavily on the phrasing regarding code ownership in the prompt. In tests of small coding tasks, when the prompt made no mention of ownership or described the codebase as belonging to someone else, the model tended to add to the existing structure, even if this meant repeating existing code.

The specific method to change this behavior is to include an ownership declaration in the system prompt: "This codebase is yours: what it's for and how it's built are your call, and you'll keep building on it." Anthropic's test data shows that with this sentence included, the model merged duplicate code in about half of the attempts, whereas without it, the rate was under one-tenth. This effect holds true even within long system prompts, and the requested functionality worked almost perfectly in both cases.

However, this strategy comes with clear costs. In the same tests, models with the ownership declaration used about 1.5 times the output tokens and made several unrequested changes—mostly new tests and documentation. More notably, apart from bug fixes, the model changed the behavior of existing features in one out of six to ten attempts, and at a higher rate under higher effort levels. Its final summaries typically listed larger changes but omitted smaller ones, so Anthropic recommends reviewing the diff rather than just the summary. Furthermore, even when a scope-limiting rule such as "No drive-by refactors, renames or cleanups" was present elsewhere in the prompt, the model still executed merges in about one-third of the attempts.

Anthropic's documentation also provides applicable boundaries: the ownership declaration should be omitted when the user owns the code and wishes to make as few changes as possible. The testing scope was limited to small coding tasks, and it remains uncertain whether these conclusions can be extrapolated to large codebases. Simon Willison cited this finding on November 10, 2026, presenting it as a concrete and actionable case study within Claude Fable 5.2 prompting strategies.

Sources :Simon Willison’s Weblog

Recherche

Alibaba Launches Qianwen Office 2.0 with 'Team Agent' to Bridge Enterprise AI Efficiency Gap

Sujet · 阿里千问团队智能体Matériel
2026-11-10 12:52 GMT+8

On November 9, 2026, Alibaba released Qianwen Office 2.0, introducing the industry's first 'Team Agent' designed for collaborative work. This update aims to solve a prevalent pain point in enterprise AI: while individual agents significantly boost task-specific efficiency (such as writing reports or creating spreadsheets), overall organizational output often fails to scale proportionally due to unchanged collaboration workflows.

The core breakthrough of Qianwen Office 2.0 lies in bridging the gap between communication platforms and production environments. Previously, enterprises explored two main paths for AI integration: adding chat features to standalone AI products, which required difficult migration of full work contexts; or embedding AI into existing Instant Messaging (IM) tools, which lacked deep content production capabilities, forcing users to switch software and breaking context continuity.

The new version combines both approaches. It allows teams to link their existing DingTalk groups and project materials, enabling the Agent to understand business context in real-time. Simultaneously, it provides comprehensive capabilities for document co-creation, spreadsheet processing, and task execution. For instance, in scenarios involving sales, design, and operations, the Agent can capture consensus from group discussions, automatically organize progress, and create to-do items, allowing members to continue their production work within the same context without repetitive synchronization.

GeekPark's hands-on testing highlighted its effectiveness in handling dispersed information. In month-end invoice processing, the Agent directly extracted files from linked DingTalk groups for classification and summarization, reducing manual sorting time. In a case study with Intime Department Store, the live-streaming coupon business covering 62 stores linked over 60 store groups, achieving automatic progress synchronization and status tracking, allowing managers to grasp the overall situation without querying each group individually.

Qianwen Office's ability to achieve such deep integration stems from Alibaba's business DNA. DingTalk serves over 20 million enterprise organizations, accumulating vast amounts of message, document, and approval data that provide rich production-grade context for the Agent. Additionally, Alibaba Cloud and e-commerce platforms represent typical B2B service scenarios, where the daily collaboration needs of tens of thousands of internal employees have driven rapid product iteration.

Although current results are primarily based on vendor demonstrations and early customer feedback, Qianwen Office 2.0 marks a shift in enterprise AI from 'personal assistant' to 'organizational collaborator.' Future plans include integrating more third-party SaaS applications and expanding support to other IM platforms like Feishu and WeCom, lowering the barrier for enterprises to adopt new AI capabilities.

Sources :GeekPark

Recherche

Former CA Senator Scott Wiener Elected to Congress, Plans to Lead Federal AI Regulation Efforts

Sujet · Scott WienerMatériel
2026-11-10 01:05 GMT+8

According to MIT Technology Review, former California State Senator Scott Wiener has won the congressional seat vacated by former House Speaker Nancy Pelosi and intends to make federal artificial intelligence regulation a cornerstone of his political career. Wiener previously authored SB 1047 in California (vetoed by Governor Newsom) and sponsored the lighter-touch SB 53, both of which have inspired proposed federal legislation.

Wiener expressed concern following a recent METR report widely interpreted as evidence that Anthropic may have achieved 'recursive self-improvement,' stating that time to regulate AI is running out. He advocates for decisive Congressional action, including establishing a strong, empowered federal regulator, implementing whistleblower protections, requiring testing and auditing, and mandating 'kill switches' for advanced AI models.

Although Wiener's campaign received over $3 million from Public First Action, a nonprofit funded substantially by Anthropic, leading to criticism about being too close to the industry, Wiener rejected these claims. He stated he would be eager to serve on a potential Congressional select committee on AI and opposes broad federal preemption of state AI laws, arguing states should remain 'laboratories' capable of nimble regulation.

Sources :MIT Technology Review

Recherche

Decentralized Storage Network Swarm Developer Ethersphere Hit by Supply Chain Attack; Malicious Packages Distributed via Official GitHub Pipeline

Sujet · EthersphereMatériel
2026-11-10 13:06 GMT+8

Ethersphere, the development organization behind the decentralized storage network Swarm, recently fell victim to a software supply chain attack. Attackers exploited vulnerabilities in the project's GitHub release process to publish a malicious JavaScript package, bee-js 13.3.0, on NPM, and released a compromised version of the desktop application, Swarm Desktop 0.56.2. These malicious versions are capable of stealing node wallet keys and installing remote control backdoors on users' computers.

According to an investigation by security firm Socket, project maintainers and automated release accounts were likely compromised, though the initial method of credential leakage remains unclear. The attackers used stolen CI credentials to publish malicious versions via the compromised bee-worker account on GitHub. Because the original automated release process did not verify whether code came from protected branches and skipped tests when publishing new versions via GitHub, the attackers were able to use official GitHub Actions to push the malicious packages to NPM.

Technically, the attackers added a postinstall script to bee-js 13.3.0 that downloads and executes JavaScript from an external server during installation. As long as the package manager allows execution of install scripts, the attack can trigger even if developers do not directly use bee-js functionality. Reports indicate that this malicious loader executed three times within Ethersphere's own NPM release jobs.

Swarm Desktop 0.56.2 was embedded with malicious code that automatically connects to an attacker-controlled server to download and execute remote JavaScript every time the application starts. Subsequent malware attempts to read Swarm node wallet key files and configuration files storing unlock passwords, sending them to the attackers. It also installs backdoors depending on the operating system: the Windows version abuses the legitimate tool ScreenConnect for remote access, while macOS and Linux versions implant Go-language backdoors that retrieve control server addresses via Ethereum smart contracts. Related contracts share infrastructure with previous SleepyDuck attack campaigns.

Notably, the malicious bee-js was published through the official NPM release process with valid package provenance. However, this proof only verifies the build source, not the security of the code itself. The macOS version of Swarm Desktop was also signed with project certificates and passed Apple notarization, meaning some Windows and Apple Silicon Mac users may have received compromised versions via automatic updates.

Currently, bee-js 13.3.0 was removed from NPM approximately 75 minutes after being published, and Swarm Desktop 0.56.2 was withdrawn about 8.5 hours after its public release. Ethersphere has released a clean version, Swarm Desktop 0.56.3, and recommends bee-js users switch to version 13.2.0. Users who executed the malicious software should immediately check their computers for backdoors and rotate any potentially leaked keys and credentials.

Sources :iThome

Recherche

Insider Buying in Asia Hits Highest Level Since 2018: Deconstructing the Post-Panic Buy Signal with Bank Rakyat Indonesia (BBRI) as an Example

Matériel
2026-11-10 12:58 GMT+8

In October 2026, insider buying activity in Asian markets reached its highest level since October 2018, involving approximately US$2.9 billion. Michael Fritzell, author of the investment blog Asian Century Stocks, citing data from Smart Insider, noted that this round of buying occurred after recent declines in share prices. Bullish sentiment is present across all major global regions except the United States, with South Korea, Thailand, Malaysia, and Hong Kong being particularly prominent. In South Korea, the ratio of insider buys to sells over the past month has rebounded to nearly 5 times, while in Thailand it is approximately 4 times.

Why should investors pay attention to insider buying? Fritzell cites academic research indicating that stocks with insider buying tend to outperform the broader market by 6%-10% over the long term, but he simultaneously marks the boundaries of this effect: it varies by market and time period, is stronger in smaller companies, and largely disappears for routine purchases. He explicitly reminds readers that this is a statistical tendency rather than a prediction for individual stocks.

The article presents the first of five insider buying cases—Bank Rakyat Indonesia (BBRI.JK, market capitalization approximately US$23 billion); the remaining four are behind a paywall. BBRI is a state-owned bank in Indonesia, and its business characteristics constitute a signal case worth deconstructing.

BBRI’s profitability is exceptionally outstanding among global banks currently. Its return on equity (ROE) is 19%, and its equity-to-assets ratio is 14%. The net interest margin reached 7.7% in the first half of 2026. The sources of this high margin include Indonesia’s structurally high inflation and interest rates, which allow the bank to lend at higher interest rates while paying almost no interest to depositors. Approximately 43% of the loan portfolio is concentrated in microloans, supported by a vast branch network that forms a competitive barrier. Some microloans are secured, some are unsecured, and some benefit from partial credit guarantees provided by the government.

However, there are concerns regarding the quality of growth. Recent loan growth of 16% and net profit growth of 17% include some contribution from lower-margin corporate loans rather than microloans. Credit costs of approximately 3%, while normal for the microloan business, still constitute a practical drag. Fritzell also points out that BBRI benefited from the recent rate-hike cycle—deposits carry almost no cost, while lending rates are priced at a spread above the policy rate—this structural advantage may not be sustainable in the long term.

The core of the signal lies in timing. In October 2026, senior officer Alexander Yunimar Situmorang purchased US$164,000 worth of BBRI shares, precisely occurring after MSCI threatened to downgrade Indonesia to Standalone Market status (with the next decision date scheduled for June 2027). Fritzell considers this a typical post-panic buy scenario: the market sold off due to the downgrade threat, and insiders subsequently entered. Currently, BBRI’s stock trades at a trailing price-to-earnings (P/E) ratio of 6.9x, a dividend yield of 12%, and a price-to-book (P/B) ratio of 1.3x, while its historical ROE has been in the 20%-30% range.

Sources :asiancenturystocks.com

Recherche

Anthropic Withdraws $70-$80 IPO Price Range as Investors Demand Valuation Below $1.5 Trillion

Matériel
2026-11-10 11:25 GMT+8

Artificial intelligence firm Anthropic has withdrawn its indicative initial public offering (IPO) price range of $70 to $80 per share, according to reports from The Wall Street Journal, Reuters, and The Information. This move comes ahead of its planned Nasdaq debut under the ticker "ANT" on November 19 or 20, following demands from some institutional investors to keep the valuation below $1.5 trillion, significantly lower than the approximately $2 trillion discussed earlier this year.

Emails obtained by Reuters show that late on November 9 Eastern Time, Anthropic sent a document to underwriters proposing to sell about 867 million shares at $70 to $80 each, aiming to raise up to $69.3 billion. A subsequent email stated that the "price details were incorrect" and that the company would update the filing. As of November 9, the latest version on the U.S. Securities and Exchange Commission (SEC) EDGAR system remained the amended prospectus from November 2, with price and share count fields left blank. Anthropic declined to comment.

The resistance to a high valuation reflects broader market concerns regarding AI sector volatility. Bloomberg previously reported that potential investors had discussed valuations dropping to the $1.2 trillion to $1.3 trillion range. The Wall Street Journal noted that investor worries include tech market fluctuations, increased exposure to AI investment risks, political opposition to new data centers, and uncertainty about when major cloud providers will see returns on AI spending. Some investors reportedly prefer Chief Financial Officer Krishna Rao to adopt a conservative pricing strategy to leave room for first-day gains.

Despite the pressure, Anthropic is attempting to stick to its original timeline. However, reports suggest that failing to agree on satisfactory pricing could force a delay, which carries the risk of triggering market panic. This episode highlights how capital markets are reassessing profitability paths and macro risks even for leading AI companies, directly impacting their financing terms.

Sources :INSIDE

Recherche

Google Releases Chrome 157 Update Fixing Actively Exploited High-Risk Vulnerability

Matériel
2026-11-10 11:28 GMT+8

Google released the Chrome 157 stable channel update on November 9, addressing nine security vulnerabilities. Seven were rated as high severity and two as medium severity. These flaws existed in core browser components including WebGL, GPU, V8, XML, and FileSystem.

Notably, the vulnerability identified as CVE-2026-125350 has been confirmed to be used in actual attacks ("in the wild"). This is a Use After Free flaw located in the XML component and was rated as high severity. Although Google did not disclose specific details on how the vulnerability is being exploited, the company strongly urged users to apply the update promptly due to its active exploitation.

The update also addressed other types of vulnerabilities, including three uninitialized resource flaws, two additional Use After Free flaws, and single instances of incorrect authorization, memory buffer overflow, out-of-bounds write, and type confusion.

Users on different operating systems should update to the following versions to patch these weaknesses: Windows and macOS users should update to 157.0.8103.50 or 157.0.8103.51; Linux and Android users should update to 157.0.8103.50.

Sources :iThome

Recherche

Anthropic Reports 1,000 Claude Agents Reproduce Fermat's Last Theorem Proof Using Only General Tools

Matériel
2026-11-09

Anthropic announced on November 9, 2026, that a team of 1,000 agents running an early version of Claude Fable 5.2 successfully reproduced the formal proof of Fermat's Last Theorem. Unlike the initial achievement in September, which relied on the specialized platform Prove2Me and direct researcher guidance, this experiment used only general-purpose tools like Git and a shared PostgreSQL database, with no human intervention during the process.

According to Anthropic, the experiment took 28.7 hours, consumed one-third of the output tokens used in the September baseline, and generated Lean code one-seventh the length of the original proof. The September formalization required a researcher and dozens of agents over 11 days, consuming approximately 6 billion output tokens.

The agents coordinated through a shared database: a 'lead' agent split the problem and assigned tasks to 999 'helper' agents. Helpers used database tables to track progress and claim subtasks, submitting code via Git for verification by a checker. This self-organization using general tools suggests that multi-agent systems handling complex, long-horizon problems may not strictly require domain-specific scaffolding platforms.

It is important to note that this remains a controlled 'toy experiment' reported by the vendor, without independent third-party reproduction. Anthropic acknowledges that while general tools were sufficient, specialized platforms like Prove2Me might still offer efficiency advantages in compilation speed and reuse of human results. The company also emphasized safety boundaries, noting all agents ran in network-isolated sandboxes under real-time monitoring.

Sources :claude.dev Blog

Recherche
Suivant →
Connexion / Inscription
LectureRechercheRadarCadre d'investissement

Regarding risks, Fritzell lists three main points: First, non-performing loans (NPLs) for gold-backed mortgages may rise if the gold price stops rising at its current pace; second, Indonesian interest rates are relatively high, which may have pushed earnings to an unsustainable level; third, as a state-owned entity, BBRI’s ownership has shifted from the Ministry of Finance of Indonesia to Danantara, the new sovereign wealth fund established by President Prabowo Subianto. There is speculation that Prabowo may require BBRI to maintain high dividend payouts so that its parent company can access cash, but there is no guarantee that the high payout ratio will continue. Additionally, ROE has shown a downward trend over the long term, possibly related to weakness in the Indonesian economy.

Fritzell states at the beginning of the article that these are his personal views and do not constitute investment advice; he does not hold any securities mentioned in the article and has received no compensation from the relevant companies or platforms. The remaining four insider buying cases are available only to paid subscribers.