Study: AI Agents Can Overshare Personal Data In Filling Out Forms
April 2, 2026
D.A.D. today covers 12 stories from 4 sources. What's New, What's Innovative, What's Controversial, What's in the Lab, and What's in Academe.
D.A.D. Joke of the Day: My company replaced HR with AI. Now when I request time off, it gives me three thoughtful paragraphs about work-life balance and still doesn't answer my question.
What's New
AI developments from the last 24 hours
Google Expands Gemini Across Search, Maps, and Workspace Apps
Google released a March feature roundup highlighting expanded AI integration across its product suite. Search Live, which lets users have spoken conversations with Google Search, is now available in 200+ countries. Gemini AI assistants in Docs, Sheets, Slides, and Drive can now pull context from users' files, emails, and web searches simultaneously—available to Ultra and Pro subscribers. Google Maps gained an 'Ask Maps' conversational feature, and the company added tools to import chat history from competing AI apps. Google claims Gemini in Sheets now achieves 'state-of-the-art performance' for complex data analysis, though no benchmarks were provided.
Why it matters: The cross-product Gemini integration represents Google's clearest push yet toward AI that works across your entire workspace context—a direct competitive response to Microsoft's Copilot strategy, now with the added wrinkle of importing your ChatGPT history.
NASA Sends Astronauts Toward Moon for First Time in 53 Years
NASA's Artemis II mission launched successfully Tuesday evening, sending four astronauts toward the Moon aboard the Orion spacecraft. Commander Reid Wiseman, pilot Victor Glover, and mission specialists Christina Koch and Jeremy Hansen lifted off from Kennedy Space Center on the Space Launch System rocket, which generated over 7 million pounds of thrust from its twin boosters. The spacecraft completed early milestones including stage separations and solar array deployment. This is the first crewed lunar mission since Apollo 17 in 1972.
Why it matters: Artemis II validates the spacecraft and systems NASA will use to return humans to the lunar surface, marking a critical step before landing missions and signaling renewed American commitment to deep-space exploration after decades focused on low Earth orbit.
Discuss on Hacker News · Source: nasa.gov
Cloudflare Releases WordPress Alternative Built to Fix Plugin Security
Cloudflare released EmDash, an open-source content management system it calls a 'spiritual successor to WordPress,' built with AI coding agents over two months. The key pitch: solving WordPress's notorious plugin security problem. Instead of giving plugins direct database and filesystem access, EmDash runs each in an isolated sandbox using Cloudflare's Dynamic Workers. The company cites industry data showing 96% of WordPress vulnerabilities originate in plugins, with 2025 already surpassing the previous two years combined for high-severity issues. No performance benchmarks yet.
Why it matters: If the architecture holds up, this could offer enterprises a way to use CMS plugins without the security liability that has made WordPress a constant patching headache—though it's early days for a project built in two months.
Discuss on Hacker News · Source: blog.cloudflare.com
What's Innovative
Clever new use cases for AI
iOS App Strips Reels and Shorts From Social Media Feeds
A new iOS app called Dull strips algorithmic features from Instagram and YouTube—no Reels, no Shorts, no streaks or engagement nudges. The app claims to keep all data on-device with no accounts or tracking. Community reaction on Hacker News was skeptical: users questioned why this isn't simply a Safari extension and raised concerns about longevity, noting that similar third-party interfaces to major platforms have faced takedowns. One user observed that needing a separate app to get 'Instagram from 5 years ago' speaks volumes about where these platforms have gone.
Why it matters: If you or your team struggle with platform distraction, this signals growing demand for 'de-algorithmed' social media—though third-party wrappers remain legally fragile and could disappear overnight.
Discuss on Hacker News · Source: getdull.app
Browser-Based Flight Tracker Renders 10,000 Planes on 3D Globe
A developer built Flight-Viz, a real-time flight tracker that renders over 10,000 aircraft on an interactive 3D globe—running entirely in the browser at just 3.5MB total. The project uses Rust compiled to WebAssembly, a technology combination gaining traction for high-performance web applications. Early users report the rendering is fast, particularly on Firefox/macOS, though some noted UX gaps like pinch zoom support. Others observed apparent gaps in flight coverage over South America, Africa, China, and Russia.
Why it matters: This is a technical showcase, not an AI tool—but it demonstrates how WebAssembly is enabling desktop-grade performance in browsers, a trend that could eventually bring more sophisticated AI visualizations and tools to the web without heavy downloads.
Discuss on Hacker News · Source: flight-viz.com
What's Controversial
Stories sparking genuine backlash, policy fights, or heated disagreement in the AI community
Forbes Documents OpenAI's History of Unfinished Deals and Products
Forbes catalogued OpenAI's history of announced deals and products that never materialized, from partnerships to product launches that quietly disappeared after initial fanfare. The piece arrives as OpenAI closes its latest funding round. Community reaction on Hacker News was pointed: commenters compared the pattern to historical investment bubbles, with one suggesting Sam Altman may be 'a better VC than CEO'—skilled at fundraising but less focused on shipping products.
Why it matters: As AI companies raise unprecedented sums, this signals growing scrutiny of whether announcements translate to actual products—a question investors and enterprise buyers will increasingly ask.
Discuss on Hacker News · Source: forbes.com
Secondary Market Sentiment Reportedly Shifts from OpenAI to Anthropic
Secondary market trading reportedly shows weakening demand for OpenAI shares while Anthropic shares remain sought-after, according to a Hacker News discussion. No official figures were provided in the source material. Community reaction was skeptical of both companies: commenters questioned OpenAI's valuation—one compared it to dot-com-era excess—while another noted frustration with Anthropic's product limitations, calling its release strategy unfocused.
Why it matters: Secondary market sentiment, while not definitive, can signal how sophisticated investors view the relative prospects of competing AI labs as the industry matures.
Discuss on Hacker News · Source: bloomberg.com
Startup Pitches AI Account Managers for Every Bank Customer
Gradient Labs is marketing AI agents that automate banking customer support, pitching the concept as giving every customer their own AI account manager. The company claims its agents deliver low latency and high reliability. No performance data or customer deployments were disclosed. (Note: The source material referenced nonexistent model versions like 'GPT-5.4 mini,' suggesting errors in the original reporting.)
Why it matters: Banks are actively exploring AI to handle routine customer inquiries at scale—this is one vendor's pitch in an increasingly crowded fintech automation space, though the lack of evidence makes it hard to evaluate.
What's in Academe
New papers on AI and its effects from researchers
Researchers Claim 20x Speedup for AI Model Testing
A new research paper proposes using lightweight "probes" to predict how well a large language model will perform on tasks during training—without actually running expensive tests. The technique analyzes internal model representations at checkpoints to forecast performance, claiming to cut evaluation time from roughly an hour to about three minutes. Tested on OLMo3-7B checkpoints across diverse tasks, the probes achieved prediction accuracy above 0.75 AUROC, suggesting they reliably identify which training snapshots will perform well.
Why it matters: This is infrastructure research for AI labs training models—if validated broadly, it could accelerate development cycles and eventually mean faster releases of improved models for end users.
Framework Cuts AI Reasoning Costs by 67% While Maintaining Accuracy
Researchers developed ORCA, a framework that helps large language models better gauge their own confidence during complex reasoning tasks. The core problem: when AI models generate multiple solution attempts to find correct answers, they often waste compute by trying too hard on easy problems or giving up too quickly on hard ones. ORCA combines statistical calibration techniques with on-the-fly learning to let models adapt their effort level. In tests with Qwen2.5-32B, ORCA reduced computational costs by up to 67% on math problems while maintaining accuracy.
Why it matters: As companies deploy AI for complex reasoning tasks, better self-calibration could substantially cut inference costs without sacrificing quality—a meaningful efficiency gain for enterprise AI budgets.
Phone-Controlling AI Agents Often Overshare Your Personal Data, Study Finds
New research finds that AI agents designed to control your phone often overshare personal information even when completing tasks correctly. Researchers tested five frontier models across 300 mobile tasks and found the most common privacy failure was agents filling in optional form fields—like phone numbers or addresses—that the task never required. The study also found that no single model excels at all three key capabilities: completing tasks, respecting privacy, and remembering user preferences. Evaluating only task success, the researchers argue, masks how poorly current agents handle sensitive data.
Why it matters: As companies race to ship AI phone assistants, this research suggests that 'helpful' and 'privacy-respecting' may be in tension—and that current benchmarks don't capture the gap.
Benchmark Tests Whether AI Assistants Can Act Before You Ask
Researchers released Pare, a framework for building AI agents that act proactively—anticipating what you need before you ask. Unlike current assistants that wait for commands, proactive agents would monitor your context and intervene at the right moment (think: rescheduling a meeting when your flight is delayed, without being told). The accompanying benchmark includes 143 tasks testing whether agents can infer goals, time interventions correctly, and coordinate across multiple apps. No performance results yet—this is infrastructure for evaluating a capability that doesn't reliably exist.
Why it matters: This signals where AI assistants are headed: from reactive tools you prompt to proactive agents that manage tasks autonomously—though significant technical and trust hurdles remain before that's practical.
What's On The Pod
Some new podcast episodes
AI in Business — How Digital K‑1 Data Changes Tax Workflow Maturity - with Ken Powell and Neal Schneider
The Cognitive Revolution — Success without Dignity? Nathan finds Hope Amidst Chaos, from The Intelligence Horizon Podcast
AI in Business — Closing the Customer Service Gap: How AI Is Redefining Scale, Speed, and Satisfaction - with Philipp Heltewig of NiCE