AI News

TurboScribe AI is a free web tool that converts audio and video recordings into editable, timestamped transcripts and subtitle files. It downloads a compact Whisper-based speech model to the browser, then performs transcription locally through WebAssembly on the device's CPU or GPU; the recording itself is not uploaded. The first use downloads and caches a model of about 39 MB, after which files are processed in 30-second pieces and the text appears progressively. Video formats including MP4, MOV, WEBM, MKV, and AVI have their audio tracks extracted automatically, while supported audio includes MP3, WAV, M4A, FLAC, OGG, AAC, and OPUS. Users can select a spoken language or use auto-detection, edit the transcript, and copy or export it as TXT, SRT, or VTT. The service requires no account, email address, credit card, or per-minute payment, but processing speed and practical file length depend on the user's hardware because there is no server-side processing limit. Long recordings are divided into ten-minute reading segments and may run more reliably on a desktop or laptop with more memory. Local processing stops if the browser tab is closed or reloaded. The compact browser model is intended to be fast and private, but it is less capable than larger server-side models on difficult audio. Background noise, room echo, overlapping speakers, strong accents, names, and numbers can reduce accuracy, so the article recommends setting the language when possible and reviewing the transcript before export. The tool lists 15 language options, with English described as the most dependable, and is positioned for meetings, interviews, lectures, podcasts, voice memos, and caption creation.

InBox Wall is a local-first Python application that places a read-only security layer between an email inbox and an AI assistant. It connects through IMAP without changing mailbox state, inlines CSS, parses HTML, and scans text nodes for hidden rendering tricks, invisible Unicode, homograph characters, and known prompt-injection phrases. The project uses inheritance-aware style analysis, a strict allow-list of visible HTML tags, background-aware checks for white-on-white text, and bounded phrase matching designed to avoid regular-expression denial-of-service risks. Clean text is stored in a local SQLite database and exposed through a REST API and Model Context Protocol server, while blocked content and original hidden text remain in a separate vault that AI tools cannot access. The design fails closed: scanner errors, invalid input, CSS inlining failures, and files over 5 MB produce a LOCKED_AWAY result with a blank body. It also limits classified text to 2,000 characters, purges quarantined originals after 30 days, and protects local files with restrictive permissions. A dashboard provides visible markers for invisible characters, threat-score explanations, review actions, and an immutable audit log. The project does not scan attachments or image text, analyze phishing URLs, or detect purely persuasive social engineering, and it is intended as defense in depth alongside existing email filters and human review. The repository includes a pytest benchmark based on whether injected tokens leak into clean text, but says its sample corpus should be expanded for production use.

Georgia Tech researchers developed ORISA, a tool that can recover fragments of deleted conversations from local AI assistants running on Android phones and rebuild the information into a working AI session. The study shows that deleting a conversation does not necessarily remove every trace when the assistant keeps information in the phone’s working memory. In one experiment, the researchers created a fictional assistant called Project Note Keeper in the ChatterUI Android app, supplied it with 100 fictional project facts, and then deleted the conversation. ORISA reconstructed a session that returned 26 facts when asked to list the notes, while the original session could actively access only 14 because other information had left its active memory. The result indicates that memory forensics may recover information an assistant was no longer actively using. The team tested ORISA across 10 Android apps and three AI models in 60 configurations, recovering information that remained in memory. Recovery becomes more difficult as later phone activity overwrites that memory. The work could aid digital investigators, but it also raises privacy concerns for users of local AI assistants. The research was presented at the 2026 IEEE Symposium on Security and Privacy.

Documents disclosed during a legal dispute involving a Canadian union suggest that Amazon used an in-house AI forecasting system called Atlas to estimate the risk of unionisation and other disruptive labour events at warehouses in the US, Canada and the UK. The documents describe Atlas as collecting hundreds of data points, including Reddit posts, responses to internal surveys, workplace information and records of previous industrial action. It reportedly assigned each warehouse a score from 100 to 900, with lower scores indicating higher perceived risk, and placed sites into five tiers ranging from no immediate risk to imminent future activity. The framework could inform staffing, resource planning and other actions intended to mitigate risks to Amazon’s labour structure. Sixteen warehouses were classified as Tier 1, including the Coventry site where worker Garfield Hylton is employed and a £500m robotic-fulfilment centre in Sutton Coldfield. The Coventry warehouse had previously seen strike action and a closely contested union-recognition ballot, which GMB lost by 28 votes. GMB, Foxglove and Leigh Day argue that the documents show intrusive monitoring and possible interference with workers’ right to organise, while continuing legal challenges also address QR codes allegedly used to help workers cancel union membership. Amazon says Atlas operates only at the site or building level, does not track individual opinions, affiliations, union membership or personal data, and is intended to identify where employees need support rather than monitor dissent. Amazon also told the publication that Atlas is not currently used in the UK, but did not say whether a similar system operates there under another name. The disclosures followed a Canadian labor-board ruling that Amazon violated local labor laws. The report also notes Amazon’s earlier €32 million French GDPR fine for excessive worker surveillance.

AI Engineering Fluency, previously called GitHub Copilot Token Tracker, is a tool for monitoring token usage and AI engineering fluency across editors, desktop applications and command-line tools. It reads local session logs, so data stays on the user’s machine unless the user enables cloud synchronization or asks the GitHub Copilot app agent to analyze the statistics. Supported integrations include GitHub Copilot in VS Code and Visual Studio, VSCodium, Cursor, JetBrains IDEs, Continue, OpenCode, Crush, Pi CLI, Claude Code, Gemini CLI, Claude Desktop, Mistral Vibe, Kiro and other listed tools. The Chromium-based VS Code extension provides real-time status-bar usage, dashboards and usage analysis, while the JetBrains and Visual Studio extensions provide corresponding tracking views. A Node.js CLI exposes usage statistics, fluency scores and environmental-impact information without requiring an editor. The project also offers a Copilot app canvas, an Oh-My-Posh prompt segment and an experimental GitHub Copilot CLI status line. Teams can deploy a self-hosted SQLite-and-Docker sharing server with GitHub authentication, optional organization restrictions and an aggregated dashboard, without requiring an Azure account. The former copilot-token-tracker extension is deprecated and prompts users to migrate.

Ozil AI is a paid video-generation tool that uses one clear, front-facing photo of a person or pet to create a 15-second recreation of Mesut Özil’s famous 2016 goal for Arsenal against Ludogorets. The generated sequence includes the chip over the goalkeeper, the touch, two sliding defenders, and the finish, while the scorer’s appearance is reconstructed from the uploaded photo. Users can keep their clothing or choose a plain football kit, select among four ByteDance Seedance models, and set 480p or 720p output, 9:16, 1:1, or 16:9 framing, plus trend music or stadium sound. Videos are purchased with non-expiring credits, starting at 10 credits for $9.99; the exact cost varies by model and quality, and rendering usually takes about five minutes. The downloaded MP4 has no watermark, and failed renders automatically return the credits. The site says photos are used only to make the video, are not used to train AI, and are not published, but users must have permission to use uploaded images and may not submit children, celebrities, footballers, or other public figures. Ozil AI is an independent fan tool with no affiliation with Özil, Arsenal, Ludogorets, UEFA, TikTok, or ByteDance, and its players and stadiums are AI-generated rather than taken from match footage.

Amazon Devices, Alexa & Leo chief Panos Panay said personal computing may move from app-centered smartphones toward ambient devices that quietly use sensors and AI to understand what users need. He made the comments at TechCrunch Disrupt alongside Amazon’s launch of new Bee devices. The product line includes a wristband that can also be worn as a pendant and a credit-card-sized device that magnetically attaches to a smartphone. Bee passively records audio and turns it into notes, to-dos, reminders and insights for later reference. The devices process audio in real time and delete the raw recordings rather than storing them, while encrypted transcripts are protected by keys created and controlled by the user’s device; Amazon employees cannot access them unless the user intentionally shares them. Panay said privacy must be designed into ambient products because they operate in the background. He expects a future in which people ask for what they need less often through individual apps because an AI layer can understand and coordinate tasks across devices. He did not predict that smartphones would disappear or that one form factor would replace another. Instead, he described connected devices with an intelligence layer that makes the surrounding world more useful.

New Jersey Deputy Attorney General Brian Richard Fitzgerald, 51, was arrested on felony charges related to allegedly creating and distributing AI-generated child sexual abuse material. Authorities say he used photos of real minors from social media to create sexualized images and shared them online. The investigation began after the National Center for Missing and Exploited Children received a tip about an illicit image shared by a Discord user in July. Detectives said digital forensics, subscriber records, and IP addresses linked the account to Fitzgerald’s home. A search allegedly uncovered prompts and scripts for generating the material, along with Discord logs in which Fitzgerald admitted to creating and sharing multiple illicit AI images. He faces charges for possession and dissemination of child sexual abuse material and criminal use of a communication facility. Fitzgerald had served as a New Jersey deputy attorney general since 2016. District Attorney Joe Khan said the state considers AI-generated CSAM as harmful as other CSAM and that the law will be enforced regardless of an accused person’s title.

The American Medical Association has released a governance framework for using agentic AI in healthcare. It says AI agents that perform clinical functions should be supported by evidence and maintain high-quality performance. The framework asks four questions: what clinical function the agent performs; what evidence shows it can perform that function and what happens if it fails; what the agent may do and under which conditions; and who is responsible for its risks and outcomes. The guidance is aimed at uses such as ambient documentation, diagnostics, and patient monitoring, where agents can complete complex, multi-step tasks with limited human oversight. The AMA warns that inaccurate or biased decisions could harm patients before a clinician can intervene, and emphasizes that physicians should remain central to development and deployment. The framework will also guide the AMA's work with other organizations. Its release comes as adoption expands: a survey of 250 U.S. healthcare leaders found that 28% had deployed agentic AI and 44% were running pilots or proof-of-concept projects, while most expected a transformative effect on clinical workflows. At the same time, 72% said such tools were sometimes implemented without IT approval. More than half identified possible compliance or regulatory violations and unauthorized access to protected health information as major concerns. The article places the framework within a regulatory gap: federal rules remain limited, state approaches vary, and Manatt counted more than 250 healthcare-related AI bills introduced in 47 states in 2025.

This field perspective argues that enterprises should treat AI governance as core infrastructure rather than a later operational layer. The author contrasts AI with traditional API programs, where fragmented patterns, ownership gaps, and lifecycle policies could often be retrofitted after systems had scaled. In LLM-based systems, decisions about data lineage, prompt ownership, model versions, access boundaries, audit trails, and compliance become embedded in the architecture from the first production deployment, while regulatory exposure can grow alongside adoption. The article also warns that retrieval-augmented generation does not remove governance problems: RAG retrieves existing organizational knowledge, so poorly cataloged or inconsistently governed data can be surfaced faster and at larger scale. It therefore presents a governed data catalog as a prerequisite for RAG rather than a follow-up task. The proposed control plane is an AI Gateway through which every AI interaction passes before reaching a model. Such a layer can detect and redact sensitive information, validate prompts against enterprise policies, enforce role- and context-based access, standardize retrieval patterns, log interactions for audit, and apply safety controls to outputs. The recommended day-one foundation includes API-first AI touchpoints, prompt versioning and ownership, model lifecycle policies, AI-specific observability, and coordination among engineering, security, legal, and compliance teams. The author concludes that governance by design should make enterprise AI more auditable and scalable than governance added after deployment.

AI engineers are becoming targets in an expanding struggle over artificial-intelligence capabilities, according to the article. The Wall Street Journal reported that OpenAI warned employees that foreign intelligence services might try to approach them through romantic or sexual “honey traps,” bribery, suspicious overtures or covert eavesdropping. One reported approach began at a San Francisco playground and turned toward the employee’s work, illustrating that exposure can occur outside controlled office environments. OpenAI’s practical guidance reportedly includes shielding laptop screens, erasing whiteboards, avoiding conspicuous company clothing and reporting suspicious contact; Anthropic and Google have reportedly issued similar warnings. The information sought would not necessarily be a complete model or blueprint: training methods, data, computing infrastructure, model weaknesses and future research priorities could still help a rival or intelligence service. The article also discusses the arrest of OpenAI employee Abdulrahman Alfozan on October 9 over an alleged concealed audio recorder found in a television USB port in an executive conference room; he was charged with unlawfully intercepting oral communications. FBI Director Kash Patel has described protecting confidential AI-development information as a counterintelligence priority, reflecting the US view of advanced AI as an economic and national-security capability. The article compares the situation with Cold War efforts to obtain nuclear secrets, while noting that AI expertise is more dispersed, researchers move between employers, papers circulate publicly and capabilities can be reproduced by teams with sufficient computing resources. Companies therefore face a difficult balance: stronger access controls and investigations are needed, but excessive restrictions could hinder recruitment and legitimate research, and nationality alone is not evidence of espionage. The article concludes that protecting AI assets requires protecting the people and institutions that create them, while distinguishing genuine threats from ordinary professional and personal relationships.

Amazon has introduced Bee Frontier, a personal AI device that can be worn on the wrist or clipped to clothing, and Bee Eclipse, a slim magnetic attachment for the back of a phone. Both devices listen when activated and turn conversations into summaries, to-dos, reminders, and recurring insights, accessible through Bee’s mobile and new desktop apps. Bee Frontier has IP67 water resistance, up to seven days of battery life, USB-C charging, dual microphones with noise filtering, and three colors. Bee Eclipse is designed to run for more than 21 days on one charge and is available in four colors. Bee can connect with Gmail, Google Calendar, Microsoft Outlook, Apple HealthKit, photos, Alexa+, and Amazon.com. These integrations can help surface action items, draft follow-up emails, create calendar invitations, provide hands-free access to Bee information through Alexa+, and offer personalized shopping recommendations. Amazon says both devices use Bee Private Compute, a secure isolated environment in which audio is processed in real time and then deleted. Audio is not stored or shared, transcripts are encrypted with keys controlled by the user’s device, and users can delete transcripts and summaries. A microphone button and indicator light show when processing is active, and users can configure automatic shutoff after periods without sound. Bee prompts users to obtain permission from call participants whenever recording starts. Bee Frontier is available now for $74.99, while Bee Eclipse will start at $99.99 when it becomes available. Basic users receive summaries, recaps, and 14-day memory at no extra cost. Bee Pro costs $11.99 per month and adds unlimited recent summaries, third-party integrations, Alexa+ and Amazon.com connections, suggested to-dos, daily memories, a personal chat assistant, templates, and the desktop app. New customers receive a 30-day Bee Pro trial.

Who Sells Data is a register maintained by Secludy AI that catalogs 161 unranked vendors serving AI labs, startups, developers, and enterprises. The directory covers reinforcement-learning environments, human and synthetic training data, evaluations and benchmarks, annotation, simulation, and RL infrastructure. Vendors can be filtered by offering, data source, task domain, and active or acquired status, then sorted by name, recent verification, or publicly reported funding. Listings span coding, computer use, enterprise workflows, finance, legal, medical, science, security, robotics, games, and other areas. Each profile is intended to link individual claims to their sources, and users can compare up to four vendors. The page also explains that SOC 2 and deployment details remain on vendor profiles and comparisons rather than serving as register-wide filters. Secludy describes its own approach as replacing real identities with stable synthetic ones across files, application state, screenshots, and trajectories to preserve training and evaluation context while removing personally identifiable and customer data.

The Calif.io team demonstrated at the Hexacon security conference in Paris what it described as the first jailbreak for iOS 27, working through iOS 27.0.1 and on the latest iPhone 18 Pro models. The jailbreak is reportedly untethered, so it remains active after a reboot and does not require the phone to be connected to a computer for reapplication. Calif.io has not disclosed which vulnerabilities make the exploit chain possible. The article says artificial intelligence helped find the chain, continuing the team’s earlier use of AI to turn an Apple fix into an exploit for a serious screen-sharing vulnerability. One possible route would be comparing iOS 27.2 beta code with iOS 27.0.1 to identify security fixes, although Apple has not published documentation confirming this. The report warns that attackers may be able to use similar prompts to build exploit tools, though the most capable security-focused models are currently limited to selected companies and researchers. Open-weight models are not subject to the same restriction. Apple’s response remains unclear, but it could issue an update after the affected vulnerabilities are identified.

Zombie AI is a paid web service that creates a 15-second fictional zombie love story video from two photos: one person is the survivor and the other becomes the zombie. The generated story follows a fixed sequence of a standoff, transformation, hesitation, embrace, and a cut to a sunny wheat-field memory, with sound included. Users can choose 16:9 or 9:16 output and pay $5.99 for 480p or $9.99 for 720p HD; there is no sign-up or subscription. The service says videos are usually ready within a few minutes and failed generations are refunded in full. It uses ByteDance’s Seedance 2.5 model and does not require users to write prompts or edit clips. Uploads are limited to clear, front-facing photos of adults who have agreed to appear; celebrities, public figures, minors, strangers, and graphic gore are excluded. Zombie AI says uploaded photos and videos are deleted within seven days, while videos remain private unless the user creates a share link.

Police arrested Chicago-area high school maintenance worker Christopher Thomas Mertens after investigators used an AI-generated image of a teenage-appearing boy named “Caleb” on the dating app Taimi. According to an affidavit by FBI Task Force Officer McKailey Scott, an unknown user moved the conversation to WhatsApp, expressed sexual interest after being told Caleb was 15, and exchanged explicit messages with the undercover investigator for several days. Investigators identified Mertens by checking his phone number with Thomson Reuters CLEAR and using Clearview AI to search images of his face. They then used Flock to locate and track his car as he drove to the planned meeting, where Livonia police arrested him. Mertens later asked to speak with Scott in jail and admitted the conduct; the complaint identified him as a high school employee in Illinois but did not name the school. The article places the case in a longer history of investigators creating fictional children online, including Terre des Hommes’ computer-generated “Sweetie” in 2013 and a more recent case involving age-regression technology. It also cites a 2025 arrest in which AI-generated child images were used in a child sexual abuse material investigation, and a 2024 New Mexico lawsuit alleging that Snapchat recommended an account using AI-generated child images to dangerous accounts. At the same time, law enforcement agencies have used AI-generated personas to monitor suspected criminals, protesters, and activists. The article notes that AI-generated child sexual abuse material is increasing and is being used to extort or entice children, while also being used by children to bully and harass peers. Livonia police did not respond to a request for comment.

Firmenza is a private company offering an “AI employee” for U.S. small businesses that use SAM.gov to register, renew registrations, and pursue federal contracts. The service can check the public SAM.gov record for an entity, monitor it daily, send reminders beginning 75 days before expiration, and guide users through renewal in SAM.gov’s order. It also matches contract opportunities by NAICS code, state, and keywords, provides change notifications and response checklists, and can match Grants.gov opportunities or New York City bids when enabled. The AI assistant can explain Firmenza, SAM.gov, registrations, and published notice fields in English or Spanish, but it does not tell users how to answer federal certifications. Firmenza says fixed rules handle public-record monitoring, most registration workflows, code finding, and contract matching; AI-generated text is used for document reading, notice explanations, and conversational guidance. Users keep control of their Login.gov account, sign in themselves, make every certification, and decide whether to bid. A Firmenza specialist can enter registration information by hand under a Data Entry role assigned by the user, and a team member can review the worksheet on request, without giving legal advice or filing on the user’s behalf. The company says it never requests Login.gov credentials and uses the public tier of SAM.gov’s Entity Management API for health checks, without showing bank or tax numbers. The health check, reminders, copilot, and worksheet review are free; specialist work is paid according to the pricing page, while the Pro package for contract matching is not yet open. Firmenza is not part of the government, SAM.gov registration and renewal remain free at sam.gov, and the company is building additional features without committed dates.