TechDaily digest, November 22, 2026
AI breakthroughs clash with security and safety concerns
This week brought remarkable AI achievements alongside serious questions about security, speed, and control. From solving century-old math problems to new restrictions on AI agents, the technology is advancing faster than safeguards can keep pace.
Budget AI services expose users to stolen credentials and data theft
Security researchers uncovered serious vulnerabilities in cheap AI inference providers, finding compromised services, stolen AWS credentials, and resellers monitoring customer requests. OpenRouter responded by removing at least three providers and over 100 endpoints from its platform.
Why it matters: Cheap AI inference services may expose users' sensitive data and API keys, creating security and privacy risks for anyone using budget-tier AI providers.
Sources: Hacker News · Hacker News · The Decoder
AI-generated proof of Riemann hypothesis verified as mathematically correct
The Lean and Mathlib communities confirmed that an AI-generated proof of the Riemann hypothesis is correct, a problem that has stumped mathematicians for over a century. The proof spans 17 million lines of code and was independently reviewed by dozens of researchers.
Why it matters: This demonstrates AI's capacity to solve foundational mathematical problems that have eluded human mathematicians for over a century, reshaping expectations for what AI can accomplish in pure research.
Sources: The Decoder
Chinese intelligence agency running covert campaign to acquire advanced AI chips
US and international officials report that China's Ministry of State Security has orchestrated a campaign to help Chinese companies obtain advanced AI chips, circumventing export controls. The espionage effort represents a significant attempt to accelerate China's AI capabilities outside official trade channels.
Why it matters: This reveals active foreign intelligence operations aimed at undermining US technology restrictions, raising national security concerns and potentially affecting global AI competition dynamics.
Sources: via Techmeme
China's tech giants doubling spending to $139B, but AI gets less than 8% of investment
China's largest technology companies plan to double capital expenditure to $139 billion in 2026, though AI labs will account for less than 8% of this spending. Analysts argue this broad infrastructure focus may be strategically more significant than Western semiconductor export restrictions.
Why it matters: This reveals how Chinese tech giants are investing in broad infrastructure rather than concentrating on AI specifically, which may affect the competitive landscape and the effectiveness of Western chip export controls.
Sources: via Techmeme
Amazon nears $7.5 billion acquisition of AI startup Decart
Amazon is in advanced negotiations to acquire AI startup Decart for approximately $7.5 billion, with a deal potentially closing within days. The acquisition would strengthen Amazon's competitive position against Nvidia in providing AI computing infrastructure.
Why it matters: Major tech companies are racing to control AI infrastructure and computing capabilities, and this acquisition signals Amazon's strategy to compete with rivals in providing AI computing resources.
Sources: via Techmeme
Major US banks ban third-party AI agents from accessing customer accounts
Major US banks are prohibiting third-party AI agents from accessing customer financial accounts and placing full liability for bot errors on users. The restrictions reflect financial institutions' concerns about security risks and errors in automated account management.
Why it matters: Users who rely on AI agents for financial management face new legal and practical obstacles, while financial institutions are signaling concerns about bot-driven errors and security risks.
Sources: Wired
AI safety testing cannot keep pace with model improvements, warns METR
METR warns that leading AI models are advancing faster than researchers can reliably test their safety, making it difficult to detect potentially dangerous capability jumps. Current evaluation methods risk both missing real hazards and overestimating what models can actually do.
Why it matters: If AI systems advance beyond our ability to test them, regulators and organizations cannot make informed decisions about deployment risks.
Sources: The Decoder
Nvidia researchers argue massive AI context windows are wasteful and unnecessary
Nvidia researchers found that large language models often do not need their full context windows, which are expensive and error-prone to run. They demonstrated that periodic summaries and keyword search can achieve similar results at significantly lower computational cost.
Why it matters: This research could make AI systems faster and cheaper to run, reducing computational costs for businesses and improving practical deployment efficiency.
Sources: The Decoder
Signal restricts mobile screenshots to block AI surveillance
Signal is limiting screenshot functionality on its mobile apps to protect users against AI surveillance and automated image analysis. The messaging app frames the restriction as a response to an existential threat posed by AI-powered visual data capture.
Why it matters: As AI tools become better at analyzing visual data, messaging platforms are taking technical steps to prevent conversations from being captured and processed without user consent.
Sources: TechCrunch
Tesla's latest autonomous driving version uses real-time AI-generated video without game engines
Tesla released Full Self-Driving v15, which features photorealistic real-time rendering of surroundings using AI video generation instead of traditional game engines. The advancement could improve autonomous vehicle safety by providing more accurate environmental perception.
Why it matters: Real-time AI video generation in autonomous vehicles could improve safety by providing more accurate environmental perception and may demonstrate new capabilities in neural rendering technology.
Sources: via Techmeme