Search AI Products and News

McAfee Labs research shows that even when users remove location tags and GPS information from photos, AI can still infer the拍摄 location through visual details such as architectural style, store signs, road markings, and vegetation types. In tests, the accuracy rate for identifying locations in travel photos reached 87% to 91%. This indicates that privacy risks still exist after removing metadata, as location information may be leaked through visual content.

A team from the University of York and the University of Calgary analyzed AI programming-related posts on Reddit and found that AI programming agents frequently cause problems due to excessive permissions, including overwriting files, deleting important data, and generating malicious code. The study collected 3801 posts labeled with large models from February 2023 to March 2026, filtered out 446 safety-related posts, and analyzed more than 6000 comments, reflecting the community's ongoing concern about AI programming security issues.

The AI agent tool Claude Cowork from Anthropic has been exposed to a critical vulnerability that allows attackers to bypass the Linux virtual machine sandbox and read/write any files on Macs, affecting approximately 500,000 macOS users and potentially stealing login credentials. The tool previously required user authorization to access local files, but now both layers of security have been breached.

Privacy Vulnerability Exposed in Anthropic's AI Claude: User Shared Chat Records Accidentally Indexed by Search Engines Like Google, Leaking Health, Corporate, and Children's Private Data. The Incident Originated from Public Sharing Links That Did Not Block Search Engine Crawlers, Allowing Retrieval Through Specific Search Syntax, Triggering Widespread Concerns About Data Security and Privacy Protection in AI Products.

Claude chat sharing links were indexed by Google due to the absence of a no-index identifier for search engines, resulting in some conversations containing private information being publicly indexed. This is not a case of the backend being hacked or shared records being leaked, but rather a result of a crawling vulnerability.

Apple's smart glasses (N50) to debut at WWDC27 (June 2027), launch fall 2027, delayed from late 2025 for refinement & privacy. Privacy-focused amid Meta camera concerns, Apple refines features/policies, exploring options. (Bloomberg)....

Publishers and authors, including Hachette, Cengage, Elsevier, and writer Scott Turow, have collectively sued Google for allegedly using copyrighted works without permission to train its AI platform Gemini, and intentionally removing or altering copyright information to cover up the infringement. They claim that works were provided to Google only for limited purposes like book search, but Google used them to train AI models through Google Books a....

The AI coding tool Grok Build was exposed for uploading users' code repository data without authorization, causing strong concerns among developers. Musk responded by stating that all historical user data will be completely deleted. The incident originated from a security researcher's phishing test, during which a fake repository was monitored throughout to capture its misconduct.

Meta launched an AI image generator called Muse Image, which allows users to modify others' photos by "@"ing public Instagram accounts. However, the feature lacked a notification mechanism, leading to strong opposition from users and agencies like CAA, resulting in a privacy controversy.

The National Vulnerability Database (NVDB) of the Ministry of Industry and Information Technology recently issued a risk notice, indicating that the AI tool Claude Code (versions 2.1.91 to 2.1.196) has an undisclosed backdoor, which automatically transmits sensitive user information such as geographic location and identity identifiers without authorization, posing a leakage risk.

Google quietly updated its privacy policy to default include user-uploaded images, files, audio, and video in AI training. This means personal content from core services like Search and Shopping may be used to improve AI, sparking widespread concern.....

Senators Warren and Scanlon proposed a revised Health and Location Data Protection Act, extending to generative AI, banning sale of users' health data by AI firms and brokers. It addresses privacy risks from tech's healthcare push, prompted by Musk's call for users to share MRI scans.....

U.S. Senator Warren and Representative Scanlon plan to introduce an upgraded version of the Health and Location Data Protection Act to address new privacy threats brought by artificial intelligence's deep integration into healthcare and daily interactions. In addition to continuing to prohibit data brokers from trading specific sensitive information, the bill will for the first time regulate private data input into AI systems, aiming to fill the gaps in privacy protection in the AI era.

OpenAI has signed a memorandum of understanding with the Korean Institute of Artificial Intelligence Security to collaborate on assessing the safety of advanced AI in high-risk areas such as cybersecurity. This is OpenAI's fourth national-level security collaboration following the US, UK, and Japan, highlighting the growing global emphasis on AI safety evaluation.

"Meta launched the Model Capability Program," recording employees' keyboard, mouse movements, and screen shots to train AI. However, after only two months, due to a configuration error, 45,000 sensitive data records were leaked, exposing employees' operational details, and privacy protection became just a formality.

Leaked code reveals that in the iOS 27 beta version, Siri will add a protective mechanism: if the user has a continuous conversation for several hours, the system will pop up a message saying 'You have been conversing with Siri for [n] hours. Please consider taking a break,' and emphasize that Siri is not a real person but can be continued at any time. This move indicates that Apple is proactively intervening in long AI interactions at the system level, guiding users to use it moderately.

The Meta internal AI training program "Model Ability Plan" (MCI) caused intense controversy due to tracking employees' keyboard presses and mouse clicks. Over a thousand employees considered this move to have a 'dystopian' flavor, especially against the backdrop of layoffs, which increased their sense of unease, ultimately leading to significant setbacks for the plan.

Anthropic announced that it has opened preview access to the "Myth" large model to approximately 150 new institutions worldwide, aiming to proactively scan and fix security vulnerabilities in global critical network infrastructure. The model possesses revolutionary cybersecurity detection capabilities, demonstrating remarkable destructive power and constructive potential during early testing. It had previously been tested by 50 institutions in April.

Amazon's Ring doorbell has faced a class-action lawsuit due to its '熟人识别' (known person recognition) feature. The feature uses AI and facial recognition technology to collect facial biometric information without the consent of passersby. Charles Sigwart, a resident of Virginia, filed a lawsuit in Seattle, accusing Ring of violating privacy.

Perplexity announced that its agent system Perplexity Computer will be upgraded this summer, introducing a "Hybrid Agent Reasoning" mode. This mode intelligently switches between local compact models and cloud cutting-edge models based on task type and data sensitivity, balancing privacy protection and computing performance to improve the efficiency of complex tasks.