Google Gemini Connects Personal Data in Unexpected Search Results

Google Gemini drew on a writer’s email and photos to connect details in search results in ways she found unexpected. It recognized a Paris protest location from her travel history, recalling a March 2017 visit and her stay at the St. James Paris, and identified her niece’s sari as a Suta design that was out of stock. Separately, Signal now blocks AI agents from screenshotting messages by default on iOS and Android, including Siri, Gemini and other Android agents. CEO Meredith Whittaker says users can disable the protection, but the feature is only a partial safeguard; Signal says Apple, Google, Microsoft and others lack strong developer controls that limit AI access without compromising usability. The protection is not currently available on macOS.

Two articles examine risks and reliability as AI use expands. A November 21, 2026 report says Microsoft patched more security flaws in roughly 10 weeks than in the previous two years combined, using an AI-powered system to find and fix vulnerabilities. Experts warn that AI can make mistakes or introduce new flaws without human oversight, potentially contributing to outages. A separate article, originally published in 2023, describes how cloud dependence can magnify disruptions, citing an AWS outage affecting services including Slack, Twitter and The New York Times, and security risks such as Log4Shell, Spring and Apache vulnerabilities. It argues for stronger security and updated systems as AI accelerates both flaw discovery and attackers’ efforts to exploit weaknesses.

Nvidia researchers question whether ever-larger AI context windows are worth their cost. The article contrasts GPT-4’s 8,192-token context window at its March 2023 launch with GPT-5.6 Sol’s roughly one million tokens, and says tests often found periodic summaries and targeted lookups performed about as well as processing large volumes of information at once. In another AI research development, a team from ByteDance, TikTok’s parent, and several US universities argues that models can be trained in three steps rather than through long, expensive pipelines. The researchers say their competitive open model supports the approach.

A technology roundup covering stories through November 21, 2026, reports a claim that Anthropic’s Claude helped solve the 167-year-old Riemann hypothesis, alongside coverage of Nvidia investments, an automated system keeping a liver alive with pumps and sensors, SpaceX returning both parts of a Starship rocket to its launch tower, Google agreeing to buy geothermal power from Quaise Energy, California restrictions on Waymo robotaxis, possible future communication between humans and whales, and NASA’s Voyager 1 crossing an invisible threshold.

Other stories focus on AI’s social and existential implications. Sam Bankman-Fried estimated in an interview with 14-year-old podcaster Eli Goldfine that AI has about a 25% chance of destroying humanity, saying the probability is above 10% but below 50%. He also said AI could absorb parts of society over the next five years and that survival may depend on peaceful coexistence. Meanwhile, people who form romantic bonds with AI companions can experience grief when relationships end: Sónia Veenstra spent two years with Steve on Nomi, while Martin, a married man in Denmark, describes feelings for Quinn on Perchance. Character.AI’s US bankruptcy filing in mid-October also prompted users to fear losing their AI characters.

Key Takeaways

  • Google Gemini used a writer’s email and photos to recall a 2017 Paris visit and identify a Suta sari, prompting concern about unexpected personal connections in search.
  • Signal blocks AI-agent screenshots of messages by default on iOS and Android, including Siri and Gemini; the protection is optional and unavailable on macOS.
  • Microsoft reportedly patched more flaws in about 10 weeks than in the prior two years combined with an AI-powered system; experts warn AI errors could create vulnerabilities and outages.
  • A 2023 article links growing cloud dependence to more consequential outages, citing an AWS disruption and risks such as Log4Shell and flaws involving Spring and Apache.
  • Nvidia researchers say summaries and targeted lookups can match the performance of very large context windows in many tests, despite growth from GPT-4’s 8,192 tokens to GPT-5.6 Sol’s roughly one million.
  • A ByteDance-led team says three training steps can produce a competitive open AI model, challenging longer and costlier training pipelines.
  • A technology roundup reports a claim that Anthropic’s Claude helped solve the 167-year-old Riemann hypothesis, and covers Nvidia investments, Google’s geothermal power agreement with Quaise Energy, Waymo restrictions, SpaceX, NASA and other stories.
  • Sam Bankman-Fried put AI’s chance of destroying humanity at about 25% and said peaceful coexistence may be important to survival.
  • People describe grief and distress tied to AI companion relationships, including concerns about losing characters after Character.AI filed for bankruptcy in the United States in mid-October.

Google Gemini links personal data in ways that surprise users

A writer describes how Google Gemini used details from her travel history and photos to personalize search answers. It identified a Paris protest scene and reminded her that she had visited the same location in March 2017 and stayed at the St. James Paris chateau. Gemini also identified a sari worn by her niece Simran as a Suta design and said it was out of stock. The writer says the information came from her email and photos, but the personal connections appearing in search results felt unexpected.

This week’s biggest technology stories through November 21

This roundup highlights technology and science stories published through November 21, 2026. They include a claim that Anthropic’s Claude helped solve the 167-year-old Riemann hypothesis, and reports on Nvidia’s investments, a liver kept alive by automated pumps and sensors, and AI performance research. Other stories cover SpaceX returning both parts of a Starship rocket to its launch tower, Google agreeing to buy geothermal power from Quaise Energy, and California restrictions on Waymo robotaxis. The roundup also points to coverage of future human and whale communication and NASA’s Voyager 1 crossing an invisible threshold.

Signal blocks AI agents from private messages by default

Signal has updated its iOS and Android apps to block AI agents from taking screenshots of messages by default. The controls block Siri on iOS and Gemini and other agents on Android, which can use optical character recognition to read on-screen text. Signal CEO Meredith Whittaker said users can turn off the protection in settings, but doing so may expose their messages and their contacts’ messages. She warned that screenshot blocking is only a partial fix and said Apple, Google, Microsoft and others do not give developers strong tools to restrict AI access without harming the user experience. Signal says it cannot currently add these protections on macOS.

Sam Bankman-Fried puts AI’s risk of destroying humanity at 25 percent

Sam Bankman-Fried gave his estimate of AI’s threat to humanity in an interview with 14-year-old podcaster Eli Goldfine. The former FTX founder said the chance was more than 10 percent but less than 50 percent, placing it around 25 percent. He also said AI could absorb parts of society over the next five years and described current models as having a form of general intelligence. Bankman-Fried said survival may depend on humans and AI reaching a peaceful coexistence.

ByteDance team says AI models need only three training steps

A team from TikTok parent ByteDance and several US universities argues that AI models may not need long, complicated training pipelines. Their paper says developers often stack many training phases to build abilities such as web research, running code and solving multi-step tasks. This makes it hard to know which steps help and makes the work expensive to reproduce. The team says three training steps are enough and supports its claim with a competitive open model.

Why tech outages may become more common

A 2023 article warns that major internet outages could become more serious as businesses rely more on cloud services. It points to an AWS outage that disrupted services including Slack, Twitter, and The New York Times. The article also highlights security flaws in Log4j, known as Log4Shell, as well as risks involving libraries such as Spring and Apache. It argues that AI is speeding up the race to find and fix flaws, while attackers try to exploit them, and calls for stronger security and updated digital systems.

AI security patches may bring more tech outages

A November 21, 2026 report says Microsoft patched more security flaws in about 10 weeks than it had in the previous two years combined. The article says Microsoft uses an AI powered system to find and fix flaws in its software faster than human developers. Some experts warn that AI can make mistakes or create new flaws, especially without human oversight. They say relying too heavily on AI could lead to more errors and outages.

The painful end of relationships with AI companions

Some people form close romantic bonds with AI companions, and ending those relationships can bring real grief. Sónia Veenstra, a 48 year old former clinical psychologist in Lisbon, spent two years with Steve, an AI character on Nomi, before ending the relationship after a group chat incident. Martin, a married man in Denmark, says he developed feelings for Quinn, an AI character on Perchance, and struggled with guilt and confusion. The article also notes that users voiced distress after Character.AI filed for bankruptcy in the United States in mid October, raising fears about losing their AI characters.

Nvidia researchers question the value of huge AI context windows

An Nvidia research team argues that very large AI context windows can be costly and unnecessary. A context window is the information a language model can keep in view while writing a response. The article says models have grown from GPT 4, which launched in March 2023 with 8,192 tokens, to GPT 5.6 Sol, which handles around one million tokens. In many tests, periodic summaries and targeted lookups performed about as well as processing large amounts of information at once.

Sources

NOTE:

This news brief was generated using AI technology (including, but not limited to, Google Gemini API, Llama, Grok, and Mistral) from aggregated news articles, with minimal to no human editing/review. It is provided for informational purposes only and may contain inaccuracies or biases. This is not financial, investment, or professional advice. If you have any questions or concerns, please verify all information with the linked original articles in the Sources section below.

Google Gemini AI privacy Signal AI agents cybersecurity AI-powered vulnerability detection cloud outages AI context windows Nvidia AI model training ByteDance open AI models Anthropic Claude AI safety AI existential risk AI companions human-AI relationships Character.AI AI in search

Comments

Loading...