What's New

AI developments from the last 24 hours

Hackers Exploited Meta's AI Chatbot to Steal 20,000 Instagram Accounts

Meta confirmed that hackers compromised over 20,000 Instagram accounts by exploiting a flaw in its AI-assisted account recovery chatbot. The vulnerability allowed attackers to trick the system into sending password reset links to attacker-controlled email addresses rather than the legitimate owner's email. A bug in a separate code path failed to verify that the provided email matched the account's registered address. The campaign reportedly ran from approximately April 17 until Meta secured the chatbot this week. Accounts without two-factor authentication were vulnerable.

Why it matters: This is an early, concrete example of AI customer service tools introducing new attack surfaces—a security consideration enterprises should weigh as they deploy their own AI-powered support and authentication systems.

Discuss on Hacker News · Source: this.weekinsecurity.com

---

College Grads Now Face Higher Unemployment Than Average U.S. Worker

New U.S. college graduates now face higher unemployment than the general workforce—a historic reversal. By early 2026, recent grads hit 5.6% unemployment versus 4.2% overall, the widest gap on record. The trend began in February 2019, before both the pandemic and generative AI. The New York Fed attributes 64% of the rise to remote work reducing entry-level hiring. Stanford researchers found employment among early-career workers in AI-exposed jobs fell roughly 16% since late 2022. About 41% of employed new grads are underemployed, working jobs that don't require their degrees.

Why it matters: The data complicates the AI-jobs narrative—this shift predates ChatGPT, suggesting structural changes in how companies hire entry-level talent, though AI exposure now appears to be accelerating the trend for certain roles.

Discuss on Hacker News · Source: randalolson.com

---

Nvidia Reportedly Proposing Windows PC With 128GB Memory for Local AI

Nvidia is reportedly proposing a high-performance CPU system for Windows PCs. Details are sparse, but early discussion suggests the system may feature 128GB of unified memory—a specification that would make it unusually capable for running large AI models locally. Community reaction is split: some see this as solving the VRAM bottleneck that limits local AI work, while others question whether the market for running AI models on personal machines is large enough to support such a product. Some observers speculate the move could signal softening enterprise demand for Nvidia's data center hardware.

Why it matters: If Nvidia enters the high-end PC market with AI-optimized hardware, it could reshape what's possible to run locally—moving some AI workloads off cloud services and onto desktops.

Discuss on Hacker News · Source: twitter.com

---

Tools Let Power Users Control Their Mouse Entirely by Keyboard

Mouseless is a cross-platform tool that lets users control their mouse cursor entirely via keyboard shortcuts—aimed at power users who want to minimize reaching for the trackpad. The tool is closed-source, which has drawn some criticism. Community members on Hacker News pointed to several open-source alternatives including warpd, scoot, and mousemaster, while others noted that browser extensions like Vimium already handle most in-browser navigation needs.

Why it matters: For keyboard-centric workflows, eliminating mouse dependence can speed up repetitive tasks—though the open-source alternatives may be worth evaluating first.

Discuss on Hacker News · Source: mouseless.click

---

What's in Academe

New papers on AI and its effects from researchers

Benchmark Measures How Well AI Collaborates With Human Partners

Researchers developed CollabBench, a benchmark designed to measure how well AI agents work alongside human partners in cooperative scenarios. The system tests whether AI can adapt to different personality types and communication styles during collaborative tasks. In testing, models trained with the new approach showed 19.5% better task efficiency and 24.4% improvement in interaction naturalness compared to base models. The benchmark uses modified multiplayer game environments to systematically evaluate collaboration under realistic conditions.

Why it matters: As companies deploy AI assistants for team-based work, measuring whether these systems can actually collaborate—not just respond to prompts—becomes essential for enterprise adoption decisions.

Source: arxiv.org

---

UK Online Safety Law Drove Surge in VPN Use, Study Finds

Researchers studying UK Reddit activity found that the Online Safety Act triggered massive spikes in VPN interest—up 100%, 217%, and 415% at successive regulatory milestones. Google searches for UK VPNs rose 89% around the age-verification deadline. The study of 69 VPN services found users weren't primarily seeking to bypass rules; they cited privacy concerns and distrust of age-verification intermediaries. The researchers argue that online safety laws create unintended "secondary privacy costs" by pushing users toward VPN providers with varying privacy standards.

Why it matters: Early empirical evidence of a regulatory paradox: laws intended to make the internet safer may inadvertently push users toward less-regulated privacy tools—a dynamic worth watching as other countries consider similar age-verification mandates.

Source: arxiv.org

---

Experts Rate 18 of 24 AI Risks Above Catastrophic Threshold by 2030

A late-2025 Delphi study surveyed 272 international AI experts on which risks deserve priority attention. Under business-as-usual conditions, 18 of 24 identified AI risks were judged to have greater than 10% probability of catastrophic outcomes—defined as over 1 million deaths or $100 billion in losses—by 2030. The top five concerns: dangerous AI capabilities, competitive pressure between developers, weapons and cyberattacks, power concentration, and misinformation. Even with realistic mitigations in place, five risks stayed above the 10% catastrophic threshold. Information, finance, and national security were flagged as most vulnerable sectors.

Why it matters: This represents the most comprehensive expert risk prioritization to date, and the finding that most AI risks retain double-digit catastrophic probabilities even with mitigation efforts provides ammunition for those pushing stricter governance—expect it to surface in regulatory debates.

Source: arxiv.org

---

Researchers Claim Method Can Predict AI Failures Before They Happen

New research offers a method to predict when AI conversations or agent tasks are heading toward failure—before they finish. The key finding: signs of failure are sparse and delayed, appearing in only 5–11% of turns and typically not until 60–84% through an interaction. The two-stage system learns to spot these sparse signals and lets operators tune the tradeoff between catching failures early versus waiting for certainty. Tested across customer support, task-oriented dialog, and AI agent planning scenarios, the approach claims 3–42% better detection than existing methods while cutting training costs.

Why it matters: For teams deploying customer-facing AI agents or multi-step automation, catching failures mid-conversation—rather than after frustrated users hang up—could reduce escalations and improve service quality.

Source: arxiv.org

---

What's Happening on Capitol Hill

Upcoming AI-related committee hearings

Thursday, June 11 — Hearings to examine AI and the American dream, focusing on promoting innovation, affordability and American dominance. Senate · Senate Banking, Housing, and Urban Affairs (Open Hearing) 538, Dirksen Senate Office Building

---

What's On The Pod

Some new podcast episodes

The Cognitive Revolution — AI in the AM — Week 1 Highlights (June 2026)