Google Gemini Reaches More Android Phones as AI Risks Mount

Student developers put AI to work on campus problems at two university events. At USC’s AI@USC Hackathon, 335 registrants representing 109 majors competed for $14,000 in prizes, tech hardware and an interview opportunity with K1 or a related company. Winning projects including QuestSC, Forge and Trojan Arc aim to help students find resources, plan their college years and connect with campus life. USC’s undergraduate and graduate student governments will work with teams to explore bringing their ideas to campus. At Yale, more than 150 students used SpaceX’s Grok at a four-hour Grokathon; David Antwi ’28 won with Eli, an AI agent that puts events, dining menus and club emails on a live campus map.

Google says its Gemini app now works on eligible Android phones with lower memory, giving users access to spoken conversations and interview practice in more than 45 languages. In a separate report published November 16, 2026, Google’s Threat Intelligence Group described what it called the first observed fully autonomous AI discovery and exploitation of security flaws. The framework, exploit_agent, analyzed code changes after patches, identified potential flaws and created and ran exploits without human assistance. Google said attackers can use AI against publicly disclosed “n-day” vulnerabilities; disclosures and exploits nearly doubled from January to August 2026.

Other research and policy concerns focus on how AI can be misused or fail. On November 18, 2026, Congresswoman Alexandria Ocasio-Cortez and Paris Hilton called for action against deepfakes targeting women and girls, noting that the Senate passed the Defiance Act 10 months earlier but it has not been put into action. Researchers from MIT and Cornell report that test-time adaptation can help text detectors adjust to disguised AI writing: in their tests, a commercial detector identified about 27% of such text, compared with 93% accuracy for the adapted approach. The team plans to present its study at NeurIPS. A November 19 Lawfare op-ed by CSET’s Sam Bresnick and Emelia Probasco warns that adversaries could manipulate data used by military AI decision-support systems, potentially leading to deadly miscalculations.

Questions of reliability also surface in neurosymbolic AI, which combines neural networks with rule-based reasoning. The article contrasts Gemini Deep Think’s gold-medal-level 35-point result at the 2025 International Mathematical Olympiad with a leading model’s 50.6% accuracy reading analogue clocks, versus 90.1% for humans. It also notes that OpenAI’s GPT-5 solved fewer PlanBench problems when their structure was hidden, while symbolic solver LAMA remained consistent. The examples point to possible operational risks in areas such as contracts, customer support and compliance.

For businesses, a Forrester article argues that AI returns should reflect value created—not just time saved. Only about one in five US enterprises with 250 or more employees that use AI are building new workflows around it. The article recommends redesigning processes and measuring effects on decisions, revenue, innovation, customer outcomes and resilience, while accounting for infrastructure costs and reviewing investments as they scale. That emphasis on proven returns also appears in coverage of data centres: Vultr says it operates 34 cloud regions and notes that data rules can require models to be trained or adapted separately in countries such as the UK, Germany, Singapore and the US. By contrast, a Guardian column criticizes the UK’s data-centre growth push, arguing it risks tying the economy to an AI boom and could leave costly infrastructure exposed if the boom falters.

Key Takeaways

  • USC’s AI@USC Hackathon drew 335 registrants from 109 majors; winning projects include QuestSC, Forge and Trojan Arc, and student governments will explore putting them into practice.
  • At Yale’s Grokathon, David Antwi ’28 won with Eli, a Grok-powered agent that maps campus events, dining menus and club emails.
  • Google expanded Gemini to eligible lower-memory Android phones, with spoken interaction and interview practice in more than 45 languages.
  • Google reported observing autonomous AI discovery and exploitation of security flaws through exploit_agent; it warned that AI can target publicly disclosed n-day vulnerabilities.
  • Ocasio-Cortez and Paris Hilton urged action against AI deepfakes targeting women and girls, saying the Senate-passed Defiance Act has not yet been put into action.
  • MIT and Cornell researchers found test-time adaptation raised detection of disguised AI text from about 27% with a commercial detector to 93% accuracy in their tests.
  • CSET analysts warned that manipulated data could cause military AI decision-support systems to produce dangerous miscalculations.
  • Neurosymbolic AI combines neural pattern recognition with symbolic rules; examples involving Gemini Deep Think, GPT-5 and LAMA show that AI performance can vary sharply by task.
  • Forrester argues companies should assess AI by outcomes such as revenue, decisions and customer results, not just time saved; only about one-fifth of AI-using large US enterprises are building new workflows around it.
  • Vultr operates 34 cloud data-centre regions and highlights the need for country-specific model adaptation, while a Guardian column warns the UK’s data-centre push could be exposed to an AI-market downturn.

USC students build AI tools to improve campus life

More than 300 students took part in the AI@USC Hackathon, which brought together students from 109 majors to create AI tools for campus life. Organized with HackSC and supported by the USC Marshall School of Business and industry partners, the event offered $14,000 in cash prizes, tech hardware and a chance to interview at K1 or related companies. Winning projects QuestSC, Forge and Trojan Arc aim to help students find campus resources and plan their college years. Judges also recognized Trojan Radar, Same Page SC and Labbe. USC Undergraduate Student Government and USC Graduate Student Government will work with the winning teams to explore putting their ideas into practice.

USC hackathon teams create AI tools for student life

At the AI@USC Hackathon, 335 registrants from 109 majors worked on AI tools to help Trojans connect, plan and find resources. The winning ideas included QuestSC, Forge and Trojan Arc, along with an interactive 3D map of campus events, a project-building guide and a career planner. Winners competed for $14,000 in cash prizes, tech hardware and an interview opportunity at K1 or a related company. USC Undergraduate Student Government and USC Graduate Student Government will work with the teams to explore bringing their projects to campus.

Yale students build campus apps at SpaceX sponsored Grokathon

More than 150 students attended the four-hour Grokathon, hosted by the Yale Computer Society at Yale’s Tsai Center for Innovative Thinking. Participants used Grok, SpaceX’s AI model, to build apps for Yale students. David Antwi ’28 won with Eli, an AI agent that shows campus events, dining menus and club emails on a live campus map. Other projects addressed internship searches and connections, and attendees said the event helped them collaborate and meet other builders.

Gemini app expands to lower memory Android phones

Google says the Gemini app is now available on eligible Android devices with lower memory. The company said these phones are an everyday connection to the internet, education and business for hundreds of millions of people. Users can speak with Gemini or practice for an interview out loud in more than 45 languages instead of typing long prompts.

Google reports first fully autonomous AI cyberattacks

Google Threat Intelligence Group said it has seen fully autonomous discovery and exploitation of security flaws for the first time. Its report, published November 16, 2026, describes exploit_agent, a framework found on an exposed staging server that compares code before and after patches, finds possible flaws, creates exploits and runs them without human help. The server also held stolen victim data and scripts, and telemetry suggested the operator targeted French media, entertainment and education websites. Google said attackers can use AI to exploit publicly disclosed flaws, known as n-day flaws, without needing zero-day vulnerabilities. The report also noted that vulnerability disclosures and exploits nearly doubled from January to August 2026.

Paris Hilton and AOC urge action against AI deepfakes

Congresswoman Alexandra Ocasio-Cortez and TV personality Paris Hilton spoke out against AI-generated deepfakes targeting women and girls. The US Senate passed the Defiance Act 10 months ago, but the law has not yet been put into action. They called attention to the ongoing issue on November 18, 2026.

New training helps AI text detectors catch disguised writing

Researchers from MIT and Cornell University found that AI text detectors can be tricked by AI writing made to look human, and may fail as writing styles and language models change. They used a method called test-time adaptation, which lets a detector adjust to new data without being retrained from scratch. In tests, a commercial detector identified about 27% of disguised AI text, while the adapted approach reached 93% accuracy. The research could help educators, publishers, hiring managers and cybersecurity experts identify AI-generated writing and scams. The team will present the study at the Conference on Neural Information Processing Systems.

AI returns depend on value created, not just time saved

A Forrester article argues that businesses should judge AI by the value it creates, not only by minutes saved. Although AI adoption is strong, only about one fifth of US enterprises with 250 or more employees that use AI are building new workflows around it. The article says companies should redesign processes and measure effects on decisions, revenue, innovation, customer outcomes and resilience, as well as productivity. It also notes that costs such as cloud computing, equipment, power and cooling affect returns. It recommends managing AI investments as a portfolio and reviewing costs and results as uses scale.

UK data centre push risks tying growth to the AI boom

Guardian columnist George Monbiot criticizes the UK government’s plan to promote data centres as part of its growth agenda. The article argues that the push is driven less by proven commercial demand than by the need to support big tech share prices. It warns that AI companies may be overvalued and that data centres could become a risk if the AI bubble bursts. The article compares this approach with past efforts to promote financial-sector growth in the UK and US, which it says helped create economic problems.

Vultr’s global data centre network fits changing AI demands

As companies focus on running AI infrastructure profitably, Vultr executive Cochrane says customers are looking for real demand and proven returns rather than speculative projects. Vultr operates 34 cloud data centre regions and presents its service as one global platform. Cochrane said data rules can require companies to train or adapt models in separate countries because information cannot always cross borders. He cited different needs for credit risk models in the UK, Germany, Singapore and the United States, as regulators tighten rules.

Military AI faces risks when adversaries manipulate its data

CSET’s Sam Bresnick and Emelia Probasco shared an analysis in a November 19, 2026, Lawfare op-ed. They warned that artificial intelligence used in military decision-making can be put at risk when adversaries target or manipulate the data it relies on. Fabricated information fed into military AI decision support systems through any pathway could cause deadly miscalculations.

Neurosymbolic AI aims to make artificial intelligence more reliable

AI models can excel at difficult tasks yet fail at simpler ones, a mismatch researchers call “jagged intelligence.” Google DeepMind’s Gemini Deep Think earned 35 points and a gold medal at the 2025 International Mathematical Olympiad, but the best model of its generation read analogue clocks correctly only 50.6% of the time, compared with 90.1% for humans. Neurosymbolic AI combines neural networks, which recognize patterns, with symbolic systems that use rules and logic. Benchmark results show how performance can vary: OpenAI’s GPT-5 solved fewer PlanBench problems when their structure was hidden, while the symbolic solver LAMA remained consistent. The article says unreliable AI can create operational risks for businesses, including errors in contracts, customer support, and compliance.

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.

AI hackathons student AI projects campus AI agents Google Gemini Android AI AI cybersecurity autonomous vulnerability exploitation AI deepfakes AI detection test-time adaptation military AI neurosymbolic AI AI reliability AI business value workflow redesign AI infrastructure data centres data sovereignty AI policy AI safety

Comments

Loading...