AI News Recap: December 26, 2025
Wall Street's Workforce Warning, Hollywood's United Front, and the Models That Keep Coming
Banks Signal Workforce Cuts, New Models Race Ahead, and Security Concerns Mount
This was the week AI stopped being theoretical for millions of workers. Major banks quietly acknowledged what many feared—those productivity gains translate directly into smaller workforces by 2026. But as Silicon Valley celebrated new model releases and faster image generation, Hollywood's biggest names drew a line in the sand, demanding consent, transparency, and a seat at the table before the industry they built gets rebuilt without them.
Note:
Firstly, I want to thank all of you who subscribe to my newsletter. Your support has been instrumental in motivating me to continue delivering informative and engaging insights on artificial intelligence to support your growth and learning. To further enhance the weekly updates, I have introduced two new sections:
1. AI Tools - which will curate practical, emerging tools to streamline workflows, boost productivity, and facilitate informed adoption amid the rapid proliferation of AI solutions.
2. Privacy / Cybersecurity - which will provide balanced updates on data privacy concerns, cybersecurity risks and vulnerabilities, regulatory developments, and best practices—equipping you to navigate AI technologies responsibly and securely in an increasingly integrated landscape.
Table of Contents
👋 Catch up on the Latest Post
🔦 In the Spotlight
🗞️ AI News
⚙️ AI Tools
🛡️ Privacy / Cybersecurity
🧩 NeuralBuddies Word Search
👋 Catch up on the latest post …
🔦 In the Spotlight
Wall Street’s AI Gains Are Here — Banks Plan for Fewer People
Category: Workforce & Skills
🏦 Major US banks including JPMorgan, Wells Fargo, Citigroup, and Goldman Sachs report significant productivity gains from AI, with JPMorgan seeing up to 6% improvement and projecting 40-50% gains in operations roles.
📉 Wells Fargo’s internal budgets indicate a smaller workforce by 2026, with the bank preparing higher severance costs for future staffing adjustments.
💻 Early AI gains are concentrated in software development (Citi reports 9% productivity improvement), customer service, operations, and regulatory reporting functions.
🗞️ AI News
Marketing Agencies Using AI in Workflows Serve More Clients
Category: Industry Applications
🎨 Marketing agencies are fine-tuning AI models on brand-specific datasets to ensure consistent visual identity, with WPP generating high-quality images for Argos in minutes instead of months.
⚡ AI compresses traditional production timelines, requiring agencies to redesign entire workflows rather than simply adding AI as a new tool layer.
🔄 New operational roles are emerging in agencies including model trainer, workflow designer, and AI governance lead, as agencies shift focus from mechanical tasks to brand stewardship.
Introducing GPT-5.2-Codex
Category: Foundational Models & Architectures
🤖 OpenAI releases GPT-5.2-Codex, an advanced agentic coding model with enhancements in context compaction, large code changes, Windows environments, and cybersecurity capabilities.
📊 The model achieves state-of-the-art results on SWE-Bench Pro with 56.4% accuracy and Terminal-Bench 2.0 with 64.0% accuracy.
🔒 It offers improved cybersecurity performance and is available in Codex surfaces for paid ChatGPT users, with API access planned and a pilot for defensive applications.
A “Scientific Sandbox” Lets Researchers Explore the Evolution of Vision Systems
Category: Research & Breakthroughs
👁️ MIT researchers developed a computational framework where embodied AI agents evolve eyes and learn to see over many generations, allowing scientists to study vision evolution.
🔬 Experiments showed that navigation tasks led to compound eyes (like insects), while object discrimination tasks produced camera-type eyes with irises and retinas.
🤖 The framework could guide the design of novel sensors and cameras for robots, drones, and wearable devices that balance performance with real-world constraints.
Gemini 3 Flash: Frontier Intelligence Built for Speed
Category: Foundational Models & Architectures
⚡ Google releases Gemini 3 Flash, a fast and cost-effective model providing Pro-grade reasoning at Flash-level speed.
📊 It outperforms Gemini 2.5 Pro on benchmarks like GPQA Diamond at 90.4% and MMMU Pro at 81.2%.
🌐 Available globally via Gemini API, app, and enterprise platforms, enabling multimodal tasks and agentic workflows.
Waymo Is Testing Gemini as an In-Car AI Assistant in Its Robotaxis
Category: Robotics & Autonomous Systems
🚗 Waymo is testing Google’s Gemini AI chatbot as an in-car assistant for robotaxis, with a 1,200+ line system prompt discovered in the mobile app code by researcher Jane Manchun Wong.
🗣️ The assistant can answer questions, control in-cabin features like temperature, lighting, and music, and use personalized greetings based on rider data and trip history.
⚙️ Gemini is instructed to maintain clear distinction between itself and the Waymo Driver autonomous system, and cannot take real-world actions like ordering food or making reservations.
OpenAI Says AI Browsers May Always Be Vulnerable to Prompt Injection Attacks
Category: Safety, Privacy & Security
⚠️ OpenAI acknowledges that prompt injection attacks on AI browsers are unlikely to ever be fully solved, as agent mode in ChatGPT Atlas expands the security threat surface.
🤖 OpenAI developed an LLM-based automated attacker trained using reinforcement learning to discover novel attack strategies internally before they are exploited in the wild.
🛡️ The company recommends users give agents specific instructions rather than broad access, and confirms Atlas is trained to get user confirmation before sending messages or making payments.
ChatGPT Launches a Year-End Review Like Spotify Wrapped
Category: Tools & Platforms
📊 OpenAI launched “Your Year with ChatGPT,” a personalized annual review feature available to free, Plus, and Pro users in the US, Canada, UK, Australia, and New Zealand.
🏆 The feature gives personalized awards based on usage patterns (such as “Creative Debugger”), creates a custom poem, and generates an image based on the user’s topics of interest.
📱 Available on web, iOS, and Android apps, users can trigger it via the home screen or by asking ChatGPT directly; Team, Enterprise, and Education accounts are excluded.
Hollywood Insiders Unite to Fight for Future of Industry With Launch of Creators Coalition on AI
Category: Ethics & Regulation
🎬 The Creators Coalition on AI (CCAI), founded by 18 industry insiders including Oscar winners Daniel Kwan and Sian Heder, launched with over 500 signatories including Cate Blanchett, Rian Johnson, and Natalie Portman.
📋 CCAI’s four core pillars include transparency and consent for content/data, job protection and transition plans, guardrails against deepfakes, and safeguarding humanity in creative processes.
🤝 The organization aims to establish shared standards and ethical practices, positioning itself as an alternative to tech companies setting terms for the entertainment industry’s AI adoption.
The New ChatGPT Images Is Here
Category: Generative AI & Creativity
🖼️ OpenAI releases ChatGPT Images powered by GPT-Image-1.5, enabling precise edits and 4x faster generation.
✏️ Supports editing types like adding, subtracting, and blending elements while preserving details.
📉 API updates reduce costs by 20% and improve handling of branded visuals for marketing and ecommerce.
AI Vs Human Code Gen Report: AI Code Creates 1.7x More Issues
Category: Testing, Evaluation & Benchmarking
📈 AI-generated pull requests contain 1.7 times more issues than human-only ones, with higher severity in logic and security.
⚠️ Logic issues are 75% more common, readability over 3 times higher, and security up to 2.74 times in AI PRs.
🛡️ Recommendations include providing context, enforcing policies, and using AI-aware checklists to mitigate risks.
⚙️ AI Tools
Chatterbox - Free Open Source Text to Speech Model
Category: Tools & Platforms
🎤 Chatterbox is an open-source TTS model family supporting multilingual, turbo mode, emotion control, and zero-shot cloning.
🏆 Outperforms ElevenLabs in blind evaluations with 63.75% preference and offers ~200ms latency.
🔓 MIT licensed, developer-friendly with pip install, available on GitHub and Hugging Face, and free for on-premise use.
Introducing Manus Design View
Category: Generative AI & Creativity
🖼️ Manus Design View enables commissioning and refining visual assets in a single AI workflow from ideas.
✏️ Supports precision editing with Mark Tool for selecting and modifying elements while preserving others.
📱 Available on mobile via Manus app for batch edits using voice or text inputs.
Building an Answering Machine
Category: Industry Applications
🦆 MotherDuck releases Answering Machine, an MCP server for natural language querying in LLMs like Claude and ChatGPT.
🔍 Agentic approach explores data, runs queries, and iterates to handle messy warehouses effectively.
📊 Enables non-technical users to gain insights, such as usage patterns and dashboard creation.
Klarna Launches Agentic Product Protocol: The Open Standard That Makes 100M+ Products Instantly Discoverable by AI Agents
Category: Tools & Platforms
🤖 Klarna launches Agentic Product Protocol for AI agents to access structured feeds of over 100 million products.
📊 Standardizes data across 12 markets for comparison, recommendations, and live pricing.
🔌 Merchants connect via API to make products discoverable without reformatting or intermediaries.
GitHub - ekzhang/jax-js: JAX in JavaScript – ML Library for the Web, Running on WebGPU & Wasm
Category: Tools & Platforms
💻 jax-js is a machine learning framework bringing JAX kernels to JavaScript on WebGPU and WebAssembly.
🔧 Supports array operations, automatic differentiation with grad(), vectorization with vmap(), and JIT compilation.
📊 Achieves high performance like 3000 GFLOP/s on benchmarks, with ONNX import and optimizers.
🛡️ Privacy / Cybersecurity
The Ghosts of WhatsApp: How GhostPairing Hijacks Accounts
Category: Safety, Privacy & Security
👻 GhostPairing exploits WhatsApp’s device-pairing to add attacker’s browser as a linked device via fake login pages.
📱 Attackers use QR codes or numeric codes to link devices, enabling message access and data harvesting.
🛡️ Prevention includes avoiding unsolicited links, enabling two-step verification, and revoking unrecognized devices.
Amazon Fends Off 1,800 Suspected DPRK IT Job Scammers
Category: Safety, Privacy & Security
🚫 Amazon prevents over 1,800 suspected DPRK operatives from joining since April 2024, with 27% quarterly increase in detections.
📄 Scammers use hijacked LinkedIn accounts, laptop farms, and fake educational backgrounds targeting AI roles.
🔍 Recommendations include identity verification, resume pattern checks, and reporting to authorities.
CISA Warns of ‘Ongoing’ Brickstorm Backdoor Attacks
Category: Safety, Privacy & Security
⚠️ CISA warns of PRC-linked actors deploying Brickstorm backdoor in VMware vSphere for government and IT sectors.
🔓 Brickstorm enables self-monitoring, encrypted C2 communications, and interactive shell access.
🛡️ Defenses include updating servers, monitoring for rogue VMs, and restricting outbound access.
Android Mobile Adware Surges in Second Half of 2025
Category: Safety, Privacy & Security
📈 Android adware detections increase 90% in second half of 2025, with PUPs up 75% and malware 20%.
🔗 Attackers chain droppers, spying modules, and banking payloads for data harvesting and fraud.
🛡️ Users advised to use official stores, scrutinize permissions, apply patches, and employ security software.
NIST and MITRE Partner to Test AI Defense Technology for Critical Infrastructure
Category: Safety, Privacy & Security
🤝 NIST partners with MITRE on AI Economic Security Center to secure critical infrastructure from cyberthreats.
🔒 Focuses on developing AI tools for protection of systems like power plants and hospitals.
📊 Emphasizes reliability, dataset protection, and assurance testing for AI in high-stakes environments.
🧩 NeuralBuddies Word Search
