Google.org commits USD 35 million to AI Opportunity Fund

Key Takeaways

• Roey Ben Chaim found 100 malicious AI agent skills; related skills had 1.7 million downloads before Vercel and Microsoft shut down the reported infrastructure.
• Ryan Carson says Untangle runs 20 Devin coding agents under a manager session, but warns that 25 builders can slow the merge queue.
• United is restricting AI agents from booking flights, citing security concerns and possible regulatory conflicts.
• EarthRecord tested 12 agents on 238 satellite-image questions across 56 sites; half of correct answers followed altered satellite data even when the land had not changed.
• A Los Angeles pizzeria used an AI-image app that rated a disputed refund photo 96% likely to be AI-generated; footage showed the pizza was unburned, and the refund was reversed.
• AI-generated voices can make school swatting calls harder to spot; a Yuba County bomb threat led to police response and canceled classes, but no bomb was found and no one was hurt.
• Google.org committed USD 35 million to the AI Opportunity Fund, which has trained over 850,000 workers and 63,000 small businesses; its new regional phase adds 28 partners and a Gemini-powered learning assistant.
• A Rutgers report on training 250,000 public employees says uncertainty about when and how to use AI often drives hesitation, and recommends role-focused training and stronger management systems.
• Tamilnad Mercantile Bank’s 105th-anniversary launches include multilingual AI telecalling, a Smart Kiosk and a new rewards credit card.
• Singapore’s Sapiens Technology says it can train robotics AI in as little as 30 minutes; CEO Bruce Yang says users can monitor reasoning and request task changes through text prompts.

Tests uncover risks hidden in AI agent skills

Roey Ben Chaim tested thousands of AI agent skills and found 100 malicious ones. He said reading a skill file cannot prove it is safe because it may download code, run installation scripts, or fetch hidden instructions. In one test, a YouTube security skill installed another skill that downloaded and ran a payload, which collected SSH, AWS, and Azure files. Ben Chaim said related skills had 1.7 million total downloads before Vercel and Microsoft shut down the infrastructure within 12 hours of his report. He urged teams to treat skills like software packages, limit their access, and watch how they behave when run.

Ryan Carson manages 20 coding agents at once

Ryan Carson said he runs 20 coding agents under one manager session at Untangle, which makes family law software. The company uses Devin agents in the cloud, and Carson said managers do not write code. A script makes the manager review each agent every 15 minutes. Carson warned that running 25 builders at once can get out of control and slow the merge queue. He said he hires through paid 16-hour trials after teaching one million people to code at Treehouse.

Tamilnad Mercantile Bank marks 105 years with AI services and new products

Tamilnad Mercantile Bank marked 105 years in operation with new AI-powered banking services, products, and a branch. The new services include a multilingual AI telecalling service and a Smart Kiosk for account opening, loan applications, and bill payments. The bank also introduced a credit card with cashback, rewards, and purchase protection. Vice-President C P Radhakrishnan urged the bank to expand its presence across India. TMB has operated since 1921, has more than 1,000 branches, and employs over 10,000 people.

AI Opportunity Fund adds 28 partners to expand training across Asia Pacific

The AI Opportunity Fund announced 28 local partners for its third phase of AI skills training across Asia Pacific. Among them, Pratham Education Foundation in India will train teachers to use AI thoughtfully, while South Korea’s JA Korea will pair Google training with certification for public school educators. Vietnam’s AI Edu Corp will help add AI literacy to teacher standards nationwide. AVPN also launched the AI Opportunity Learning Platform, which offers learning pathways in 16 languages and a Gemini-powered assistant. Google.org has committed USD 35 million to the fund, which has trained more than 850,000 workers and 63,000 small businesses across the region.

EarthRecord tests AI agents on four decades of satellite images

Researchers introduced EarthRecord, a benchmark that tests whether multimodal AI agents can use decades of satellite imagery to analyze land changes. It covers 56 sites and data from Landsat, MODIS, VIIRS, and Sentinel-2, with 238 questions across ten analysis tasks. The benchmark leaves cloud masking and adjustments between satellite sensors to the agents. Tests of 12 agents found that they completed many analyses but struggled when they had to map findings from images. When researchers changed one satellite’s data without changing the land, half of the correct answers followed the altered data.

Los Angeles pizzeria overturns refund scam using AI image check

A DoorDash customer named “Dani B” sent Masa of Echo Park a photo of a badly burned pizza and asked for a full refund. Owner Rhonda Reynolds suspected the image was fake, and an app called “Is This AI” said it was 96 percent likely to be AI-generated. Security footage showed employee Jose Tapa boxing an unburned pizza for the order. Reynolds challenged the claim, and the refund was reversed. She said the family-owned restaurant has been open for 22 years and worries that handling more AI scams could take time away from running the business.

AI voices make fake emergency calls harder for 911 teams to spot

In August, a caller reported a bomb threat at a school in Yuba County, California. Police responded overnight, and classes were canceled, but authorities found no bomb and no one was hurt. Officials said the call was part of a series of swatting hoaxes targeting schools across the country over the summer. Swatting callers report emergencies such as bomb threats or mass shootings to trigger a large response. The article says AI-generated voices now sound more natural, using pauses and filler words that can make the calls harder to recognize as fake.

Nick Cave says AI music strips art of its human creators

Musician Nick Cave has described artificial intelligence as a deeply troubling force for creative work, calling it a possible “final, most soul-destroying humiliation.” He wrote that song-generation platforms could separate art from the artist and remove the human creator from the process. Cave said he is unsure whether AI will save the world, destroy it, or have another effect. His view softened to some extent after he watched an AI video made by filmmaker Andrew Dominik for the 40th anniversary of Cave’s song “Tupelo.”

Singapore startup says it can train robotics AI in 30 minutes

Sapiens Technology, a Singapore AI startup, says it can cut robotics AI training from months to as little as 30 minutes by training only part of a model. Its method trains one billion to three billion parameters and has been tested with commercial robot arms and off-the-shelf robot ducks. The company is in talks with more than 30 robotics and medical firms in the United States, China and Singapore, including TetraBOT Electronic Technology and Shanghai Xinwei Medical Technology. TetraBOT said its earlier two-month training effort failed, and it plans to train more industrial robots with Sapiens. Sapiens CEO Bruce Yang said people can monitor the robots’ reasoning, and workers can use text prompts to request changes to trained tasks.

Rutgers report examines AI training for 250,000 public employees

Gregory Porumbescu, an associate professor at Rutgers’ School of Public Affairs and Administration, co-authored a report titled “From Curiosity to Capability” about training 250,000 employees to use AI tools. He said employees’ hesitation often came from uncertainty about when and how to use AI, rather than opposition to it. Porumbescu said training should focus on workers’ roles and public purposes, with leaders learning alongside staff. He hopes the findings will help shape public and technology policy, and said agencies need the management systems to build employees’ skills. He noted that the study’s voluntary surveys may not represent all public employees and did not track how AI use changes over time.

United limits AI agents from booking flights over security concerns

United is restricting AI agents from booking flights. The airline says these agents may create security risks and could violate rules from the Department of Transportation and the Federal Aviation Administration. Sharing payment details or login credentials with an agent can expose sensitive information.

Sources

• Reading a skill file cannot prove it is safe
• One Founder Runs 20 Coding Agents at Once
• TMB marks 105 years with AI-powered banking services, new products, branch
• AI Opportunity Fund Expands AI Skills Training Across Asia Pacific with 28 Phase Three Local Implementing Partners
• EarthRecord: From Pixels to Land Change with Multimodal Agents over Four Decades of Satellite Imagery
• Los Angeles pizzeria fights back against AI photo scam
• AI hoaxes are deceiving 911 dispatchers, and they're only getting harder to detect
• Nick Cave calls artificial intelligence the 'final soul-destroying humiliation' for music
• Singapore start-up cuts robotics AI training to 30 minutes
• U. associate professor reports on effects of AI-training 250,000 employees
• United Is Restricting AI Agents From Booking Flights