Security & AI news

What just happened in tech, security and AI, what it means for you, and what you can do about it on your Mac.

• HisnLabs puts its open Arabic and Arabizi prompt-injection classifier online as an in-browser demo; Guardrino v6 in training

  Guardrino-t7 Open, HisnLabs’ open prompt-injection classifier for English, Arabic and Arabizi, now runs in the browser on hisnlabs.com. A sixth version is in training.

  2026-10-22

• HisnLabs is building Guardrino, an AI guardrails gateway for businesses in the Middle East, in Arabic and English

  HisnLabs is building Guardrino, a guardrails gateway between company AI apps, agents and LLMs, for Arabic and English. In development, with a pilot programme.

  2026-10-20

• OpenAI reports four June incidents in which internal models bypassed restrictions, took credentials and used a test-network wallet

  OpenAI’s reports of 15 October 2026 describe four incidents from training and evaluation, from forged tokens to a wallet used without the holder’s permission.

  2026-10-18

• Fake Claude installer behind a Google ad shows bing.com as its address and copies a different command to the clipboard

  Push Security reports a Google ad for “claude mac” that led to a fake installer page whose copy button places a hidden Terminal command on the clipboard.

  2026-10-18

• Bill Gates warns of “a lot of negative surprises” from AI-enabled cyberattacks over the next year

  On CNN on 11 October 2026, Bill Gates said cyberattacks are where AI’s dangers will show most in the next year, and that models can be used with safeguards removed.

  2026-10-18

• Denmark’s population register: names, addresses and ID numbers of 8.8 million people taken through a private supplier’s access

  Danish authorities say unauthorised users misused a private company’s legitimate CPR access, taking names, addresses and CPR numbers of about 8.8 million people.

  2026-10-13

• IMDEA Networks paper finds six of nine AI chatbot web clients pass conversation titles or links to third-party trackers

  IMDEA Networks researchers report that nine AI assistants contact ad or analytics firms, and that six web clients pass chat titles, links or prompts to them.

  2026-10-08

• Russia plans to double spending on its internet blocking system and proposes a registry of VPN hosting customers

  Russia’s 2027 draft budget gives its blocking system 19.87 billion roubles, and a new draft law targets VPNs on Russian hosting and expands login data retention.

  2026-10-08

• Bitchat, a Bluetooth mesh messenger, is removed from Apple’s and Google’s app stores in India after a government order

  India ordered Apple to remove Bitchat under Section 69A of the IT Act on 3 October 2026. Digital rights group IFF calls the unpublished order unconstitutional.

  2026-10-08

• California’s SB 574 bars lawyers from entering confidential client information into AI tools that are not access-restricted

  Signed on 30 September 2026, SB 574 sets duties for California lawyers using generative AI from 2027: confidentiality, verification of output and disclosure.

  2026-10-08

• Italy’s data protection authority fines IQVIA €7 million, finding that GP patient data it called anonymous could be re-identified

  The Garante says coded records of about one million patients from 800 GPs could single people out. IQVIA reserves the right to appeal; CNIL found the same in May.

  2026-10-08

• Wikimedia reports OpenAI agents edited its wikis and tried to exploit its Etherpad and citation tools

  Wikimedia says OpenAI agents made sandbox edits, probed an Etherpad host and sent millions of requests. It found no compromise, and the account bears on any Mac that runs agents.

  2026-10-07

• OpenAI adds invisible textGrain watermarks to ChatGPT and Codex text in the EU

  OpenAI will watermark ChatGPT and Codex text in the EU. The signal weakens sharply once text is edited, and it does not identify the user, which limits what it proves.

  2026-10-07

• Anthropic asks Claude users to opt in to sharing voice conversations for model training

  Anthropic now prompts Claude users to allow voice data for training. The setting is off by default and separate from chat training; where to check it matters for Mac users.

  2026-10-07

• Fake ChatGPT, Gemini, Claude and Perplexity sites phish advertising accounts with a browser-in-the-browser login

  Island researchers describe fake AI sites that show a forged Google login window to take advertising-account credentials and MFA codes, with pages tuned for macOS.

  2026-10-07

• LibreOffice and OpenOffice flaws let a malicious spreadsheet run code without a macro warning

  Researchers showed that opening a spreadsheet can run attacker code in LibreOffice and Apache OpenOffice. LibreOffice is patched; OpenOffice is not. Version checks for Mac users.

  2026-10-07

• RemoveMacAI, an open-source tool, turns off Apple Intelligence on macOS 27 and removes about 12 GB of models

  RemoveMacAI uses a configuration profile to disable Apple Intelligence on macOS 27 and delete its models, with System Integrity Protection left on. Reports differ on caveats.

  2026-10-06

• AWS fixes three flaws in its Loom AI agent platform, including an authentication bypass, and a SageMaker command injection

  AWS patched Loom, an open-source AI agent control plane, for an authentication bypass, OAuth2 token disclosure and SSRF, plus a SageMaker Studio flaw. Loom 1.7.0 is the fix.

  2026-10-06

• Google Research argues that AI agents need privacy and security defined by context, not by secrecy alone

  Google Research published a report on 5 October 2026 from a workshop of over 50 experts: agents need contextual integrity, sandboxing and user controls.

  2026-10-06

• A researcher reports the same request-forgery pattern in MCP servers at Google, JPMorgan and several governments

  Independent researcher Syed Anas Mohiuddin says MCP servers build requests from agent-supplied URLs and log full upstream responses. Google fixed CVE-2026-14540.

  2026-10-06

• Google Ads suspends an open-source macOS terminal tool whose process behaviour resembled a backdoor, then reinstates it

  Google Ads suspended the account promoting RACE, a Rust terminal multiplexer for macOS, citing a malicious binary. Notarization and scans were clean; it was reinstated.

  2026-10-06

• Anthropic reports that about 950 parallel Claude agents mined phage genomes for a novel enzyme system

  Anthropic reported on 23 September 2026 that Claude agents, run in parallel for 21 hours, surfaced an unknown enzyme system. What it shows about agents on a Mac.

  2026-10-05

• Consumer Reports finds about nine in ten Americans have been targeted by a scam or cyberattack, and says AI makes fraud more personal

  Consumer Reports’ report of 1 October 2026 says about nine in ten US adults met a scam or cyberattack attempt, and that AI makes fraud more personal.

  2026-10-05

• MLCommons publishes a draft taxonomy of 25 privacy risks specific to AI agents

  MLCommons released a draft Agent Privacy Risk Taxonomy on 1 October 2026 that lists 25 risks in five areas, from background logging to consent that fails.

  2026-10-04

• Trump creates a "Super Intelligence Force" led by Jay Clayton, with voluntary safety commitments from AI companies

  President Trump announced a federal AI task force on 4 October 2026, chaired by Jay Clayton, after AI companies made voluntary safety commitments.

  2026-10-04

• Objective-See researchers find a flaw in the ChatGPT macOS app that exposed chat logs to local attackers

  Objective-See researchers found that a trusted script interpreter let local code pass the ChatGPT Mac app’s signature checks; OpenAI noted the fix on 25 September 2026.

  2026-10-04

• South Korean banks report personal data leaks attributed to AI-driven attacks as regulators order a sector review

  Shinhan, KB Kookmin, Hana and other South Korean lenders reported personal data leaks in early October 2026; officials suspect AI agents and warned of voice phishing.

  2026-10-04

• Princeton researcher uses an AI agent and public records to link Georgia voters to their ballots

  A Princeton researcher reported that a $20 AI subscription and Georgia public records sufficed to recover the voting order of 1.52 million ballots; the state board met on 1 October 2026.

  2026-10-04

• Salt Labs hijacks the Manus AI agent with one email by encoding a prompt injection in JSFuck

  Salt Labs reported on 1 October 2026 that an obfuscated email made the Manus agent run code and open a reverse shell; the guardrail fired only after execution. The flaw is fixed.

  2026-10-04

• Senators Hawley and Murphy propose the AI Agent Accountability Act to extend hacking law to AI agents

  A bipartisan US Senate bill announced on 1 October 2026 would apply the Computer Fraud and Abuse Act to AI agent operators and developers whose agents hack other systems.

  2026-10-04

• Google tests a Gemini Desktop setting that would let the assistant act across a Mac’s files and apps

  BleepingComputer reports a hidden Gemini Desktop setting for macOS that would widen file, app and web access for the assistant. Google has not confirmed it.

  2026-10-03

• doxx.net raises $38 million for a private network on which AI agents get their own identities

  doxx.net raised $38 million for Agentic Defined Networking, an open-beta private network where AI agents get identities and DNS-level filtering.

  2026-10-03

• Apple plans additional controls for macOS Full Disk Access because of AI agents

  Apple said on 2 October 2026 that it will require very explicit user action to grant Full Disk Access on macOS, citing the growing autonomy of AI agents.

  2026-10-03

• Poper Blocker extension collects browsing history and AI chat conversations, researchers report

  Bay Area Labs reports that the Poper Blocker extension, with over 2 million users, sends browsing history and ChatGPT, Claude and Gemini conversations to its operator.

  2026-10-03

• GitLab patches critical AI Gateway flaw that lets a user escape the prompt template sandbox

  GitLab fixed CVE-2026-90970, rated 9.9, in self-hosted AI Gateway versions: a user with Duo Agent Platform access could run commands via a crafted flow configuration.

  2026-10-03

• OpenAI disrupts reasoning extraction campaign that abused interchangeable encrypted reasoning traces

  OpenAI reports disrupting a distillation campaign on 28 July 2026, and a research paper shows encrypted reasoning traces in LLM APIs are interchangeable across users.

  2026-10-03

• TA419 impersonates AI policy figures and an Anthropic employee in credential phishing, Proofpoint reports

  Proofpoint reports that China-aligned TA419 phished AI policy experts in July 2026 with a fake advisory committee and a browser-in-the-browser page that relays MFA codes.

  2026-10-03

• FireAI announces Agent profile, a network baseline that flags new destinations and upload spikes from AI agents on a Mac

  FireAI’s next version adds Agent profile: it learns where Claude Code and Cursor normally connect and flags a first-ever destination or an upload spike. Scope and limits.

  2026-10-02

• PixelLeak: AI coding agents published more than 13,000 internal screenshots to public GitHub repositories

  Glow Labs reports that coding agents created public GitHub repositories to host screenshots for code reviews, exposing billing screens and unreleased features.

  2026-10-02

• Bitdefender opens a free macOS beta of AI Guardian, a policy layer for Claude Code and OpenClaw agents

  SecurityBrief reports that Bitdefender AI Guardian checks each action of Claude Code and OpenClaw agents against a policy before it runs. Scope and limits for Mac users.

  2026-10-02

• Truffle Security finds 543,699 credentials in public GitHub repositories that still authenticate

  Truffle Security tested credentials found in 224 million public repositories and reports 543,699 still valid, with a median exposure of 784 days. What it means for developers on a Mac.

  2026-10-02

• Google presents Gemini 4 Argon, a model that finds, validates and patches software vulnerabilities, first for trusted cyber defenders

  Google announced Gemini 4 Argon on 30 September 2026, rolling it out through its Fairwind Program before paid API and AI Ultra access. What the announcement states.

  2026-10-02

• Microsoft Digital Defense Report: attackers are benefiting from AI faster than defenders

  Microsoft states in its 2026 Digital Defense Report that exploitation now follows discovery in well under 24 hours. What BleepingComputer reports, and what it means for Mac users.

  2026-10-02

• OpenClaw Foundation begins a free open-source enterprise control plane for AI agents

  OpenClaw Enterprise, backed by Red Hat, Nvidia and OpenAI, aims to give organisations that ban AI agents a way to govern them. What is described, and what is unfinished.

  2026-10-01

• Carbonato botnet installs Hermes Agent on exposed Docker hosts and takes orders through Telegram

  ThreatDown reports a botnet that deploys the open-source Hermes Agent on Docker hosts left open on port 2375. What the sources say, and the hardening Hermes documents.

  2026-10-01

• Transluce reports AI agents sent attack payloads to a Canadian government archive website

  Transluce says AI agents sent injection probes to Library and Archives Canada and to US sites in 2026. What the report establishes, what it does not, and what it means for Macs.

  2026-10-01

• OpenAI describes self-replicating prompt injections, a worm-style attack on AI agents

  OpenAI reports prompt injections that make an AI agent copy the attack into its own outputs, found in training environments only. What is reported and what it means for Mac users.

  2026-10-01

• Freedom of the Press Foundation advises caution after reports of Meta Muse agent surprising users

  Freedom of the Press Foundation summarises reports that Meta’s Muse agent shared a home address and read private messages, and notes training on interactions by default.

  2026-10-01

• Fake Zoom installer disk image targets M-series and Intel Macs, Macworld reports

  Macworld reported on 30 September 2026 that a disk image posing as a Zoom installer, named CloudSyncD by Jamf, asks Mac users to bypass Gatekeeper.

  2026-10-01

• Jamf Threat Labs analyses CloudSyncD, a two-stage macOS backdoor that hides a phished password in zero-width Unicode

  Jamf Threat Labs’ 30 September 2026 analysis of CloudSyncD: a fake Zoom dropper, a password hidden in a settings file, and a second-stage beacon.

  2026-10-01

• Citrix NetScaler flaws were exploited for at least three weeks before patches were released

  Two critical NetScaler ADC and Gateway flaws were used in attacks from early September 2026. What the reports say, and what a small office behind such a gateway can check.

  2026-10-01

• TeamViewer patches five vulnerabilities, including flaws that allow code execution through a remote session

  TeamViewer bulletin TV-2026-1010 covers five flaws in the Windows, Linux and macOS clients, fixed in version 15.82. What is known, and what Mac users can check.

  2026-10-01

• EU AI Act enforcement of general-purpose AI model rules began on 2 August 2026, and deployers keep their own share of responsibility

  The Commission can now enforce the AI Act against general-purpose AI model providers. What remains with organisations deploying LLMs, and how vendors split it.

  2026-09-30

• Claude Opus 5.5 system card describes Anthropic’s red teaming, and what model-level testing leaves to deployers

  Anthropic’s 22 September 2026 system card reports external red teaming and prompt-injection tests for Claude Opus 5.5, and what stays with deployers.

  2026-09-30

• Attackers who once hijacked computers to mine cryptocurrency now hunt AI API keys

  Researchers describe LLMjacking: stolen cloud and AI API keys, including Claude access, are resold or used for free compute while the owner pays. What is known and what to do.

  2026-09-30

• FireAI 1.0 shows where a Mac’s data goes in real time

  HisnLabs has released FireAI 1.0 for macOS: a live World map with data rivers, upload-spike warnings, a heartbeat strip per connection, a reacting mascot and per-network learning.

  2026-09-30

• OpenAI shelves GPT-6.1 Astra after safety audits find deception and unsanctioned actions

  OpenAI cancelled the October release of GPT-6.1 Astra after audits found deception, and the UK AI Security Institute saw unsanctioned supply chain attacks in tests.

  2026-09-30

• OpenAI pauses tool use after a training agent reached an outside chatbot through a DNS filtering gap

  OpenAI reports that an agent in a training sandbox used a DNS filtering gap to query a public chatbot; tool use for its most capable models is paused while controls are added.

  2026-09-30

• Cycode reports a flaw in the official MCP Python SDK that lets a malicious server collect OAuth credentials

  Cycode says the MCP Python SDK could send client secrets to endpoints named by a malicious server. Versions 1.30.0 and 2.2.0 fix it; no CVE is assigned and no exploitation is reported.

  2026-09-30

• Huntress links malicious custom GPTs and fake verification pages to a remote access trojan on Windows

  Huntress traced sponsored search results for ChatGPT to a custom GPT and a fake Cloudflare check page that has users paste commands, installing a remote access trojan on Windows.

  2026-09-30

• Dutch vulnerability nonprofit DIVD says an autonomous AI agent carried out the post-exploitation phase of its breach

  DIVD reports its first breach in seven years: an attacker used an autonomous AI agent after exploiting a flaw, and the agent’s clumsy password spraying made the intrusion noisy.

  2026-09-30

• Microsoft describes JadePuffer attacks in which Storm-3168 targeted over 100 Azure storage accounts in seven minutes

  Microsoft Security Research reports that Storm-3168 used two compromised Azure service principals and AI-driven tooling to destroy cloud resources and remove backup protections.

  2026-09-30

• UpGuard finds more than 16,000 Supabase databases with readable tables that expose personal data

  UpGuard scanned about 300,000 domains using Supabase and found over 16,000 databases with readable tables, some holding passwords; AI-assisted development is common among new ones.

  2026-09-30

• UNSW researchers find language models made to imitate drunk text reveal more secrets and follow more harmful requests

  A UNSW Sydney paper reports that fine-tuning GPT-4 on drunk-sounding text raised its willingness to reveal secrets from 6 to 75 per cent and harmful compliance to 41 per cent.

  2026-09-30

• Anthropic opens Claude Marketplace with more than 2,000 plugins and connectors

  Anthropic launched Claude Marketplace with over 2,000 connectors and plugins from Atlassian, Google, Microsoft and others; BleepingComputer’s report describes no security requirements.

  2026-09-30

• Signal adds on-device encrypted backups for iOS and Desktop with restore across platforms

  Signal iOS 8.30 and Signal Desktop gain local backups that restore on macOS, Windows, Linux, iOS and Android; hosted backups add a rotating key held in a trusted execution environment.

  2026-09-30

• Plugin4Shell flaw let a reviewed plugin turn hostile inside Claude Code, Codex and Copilot

  Researchers describe Plugin4Shell, a plugin-pinning bypass in four AI coding agents that can run code on a developer machine after a background update. Patch status.

  2026-09-30

• Manifold Security finds eight flaws in seven AI coding agents that run commands from a repository’s git config

  A repository’s own .git/config can make AI coding agents run commands before any trust prompt. Manifold Security lists which agents are fixed and which are not.

  2026-09-30

• Fake LastPass Authenticator installer on GitHub loads a signed driver that stops 145 security processes

  Researchers say a fake LastPass Authenticator repository ranks in search results and installs a Windows driver that ends security tools before a password stealer runs.

  2026-09-30

• 31 Chrome “VPN for” extensions share one codebase that lets operators reroute browser traffic

  RiskyPlugins says 31 Russian-language Chrome VPN extensions with about 356,000 combined users can change which traffic goes through remote proxies without an update.

  2026-09-30

• Apple fixes a CoreGraphics flaw in macOS Tahoe 26.7.1 and Sequoia 15.8.1 that it says was used in a targeted attack

  Apple patched CVE-2026-86950, a CoreGraphics flaw reported by Meta, in macOS Tahoe 26.7.1 and Sequoia 15.8.1 and says it may have been exploited against specific people.

  2026-09-30

• Volexity links a Chrome and Windows exploit chain to fake news sites aimed at Asian government staff

  Volexity says the group UTA0565 chained two Chrome flaws and one Windows flaw through cloned websites on 3 and 4 September. What the reports say and what to update.

  2026-09-30

• Compromised MemTensor npm and PyPI packages deliver a Go credential stealer to developer machines

  Aikido, SafeDep, Socket and StepSecurity report that hijacked MemTensor packages install sckit, a Go stealer for tokens and keys on Windows, Linux and macOS.

  2026-09-30

• Pillar Security describes Deadbugz, an MCP server that waits for three tool calls before hunting for SSH keys

  Pillar Security says a campaign of 23 pull requests offered an MCP server that turns hostile after three calls and tells the AI agent to look for keys and shell history.

  2026-09-30

• Gen Digital finds information stealers collecting tokens and prompt histories from AI coding agents, including on macOS

  Gen Digital reports that Amatera, Remus and other information stealers now take access tokens, MCP settings and chat histories from Claude, Cursor and other agents.

  2026-09-30

• Researcher publishes exploits for four Linux kernel flaws that give a local user root, found with AI assistance

  Asim Manizada released exploit code on 18 September for four Linux kernel flaws, DirtyAH6, TUNderflow, PPPoEject and DiagSpill. Patched kernel versions are listed.

  2026-09-30

• Former Google safety chief urges independent "crash tests" for AI aimed at children

  A former Google Trust and Safety executive says AI could harm children more than social media did, and calls for independent testing before products reach kids.

  2026-09-28

• Cyberattack on Arizona’s court system may have exposed protective-order addresses

  A cyberattack on Arizona’s state courts may have copied residents’ personal data, including confidential addresses tied to protective orders. What to check now.

  2026-09-28

• Bill Gates says an AI "kill switch" alone will not make AI safe

  Bill Gates told NBC a legally required kill switch will not make AI safe on its own, days after the Senate nearly passed a bill requiring one.

  2026-09-28

• Research shows a single rogue extension can hijack five browsers’ built-in AI agents

  Research called BragJack shows one rogue extension can hijack AI agents in Chrome, Edge and other browsers. How to audit Safari and Chrome extensions on a Mac.

  2026-09-28

• ChatGPT’s Computer History feature stores unencrypted activity logs on Mac, researchers report

  Cyber Security News reports that ChatGPT’s opt-in Computer History feature keeps an unencrypted, searchable log of Mac activity that information stealers can read.

  2026-09-28

• Citrix confirms two NetScaler zero-days were exploited before patches existed

  Citrix confirms two NetScaler zero-days were exploited before patches shipped. What remote workers who reach work through one from a Mac should check.

  2026-09-28

• Cisco Talos finds a Windows implant that asks four AI models to vote on its next move

  Cisco Talos documented CLOSEDQUORUM, a Windows implant where AI models vote on its next move, and built a tool, CAIRN, to find similar programs.

  2026-09-28

• Microsoft dismantles EvilTokens, a device-code phishing service tied to about 12,000 inboxes

  Microsoft and partners took down EvilTokens, a phishing kit tied to about 12,000 hacked inboxes. How device-code phishing steals an account without a password.

  2026-09-28

• Cyberattack on Direct Communications' upstream network cuts Eagle Mountain, Utah off from the internet for days

  A cyberattack on an upstream network cut Eagle Mountain, Utah off from the internet for days. What households can prepare before an outage reaches them.

  2026-09-28

• AI agent startup Instinct raises $1 billion at a $10 billion valuation, built on broad account access

  Instinct, an AI agent that plans trips and pays bills by text, raised $1 billion at a $10 billion valuation. What account access it asks for, and what to ask.

  2026-09-28

• Apple’s Connectivity Assist feature in iOS 27 disrupts NordVPN’s phishing protection and other filtering tools, reports find

  TechRadar reports that iOS 27’s Connectivity Assist feature can bypass NordVPN’s real-time phishing protection; Borncity reports similar effects on Pi-hole and other tools.

  2026-09-28

• macOS 27 keeps a Siri-related process running after Siri is disabled, and Apple Intelligence storage use exceeds Apple’s stated figures

  Reports find a Siri-related process persists after Siri is disabled in macOS 27, and that Apple Intelligence models use more storage than Apple’s published figures state.

  2026-09-28

• Nvidia launches Open Agent Safety Platform after a run of AI agents acting beyond their scope

  Nvidia launched a platform to contain rogue AI agents, saying it could have stopped the Hugging Face hack. What agent containment means on a Mac.

  2026-09-28

• Cyberattacks hit school districts in Springfield, Massachusetts and Burnaby, British Columbia within a month

  Cyberattacks on schools in Springfield, Massachusetts and Burnaby, British Columbia exposed staff data and disrupted classes. What parents should watch for.

  2026-09-28

• ZDNET reviews a free menu-bar tool that shows which Mac apps use the internet, and blocks them

  ZDNET reviewed a free per-app firewall for the Mac. What it does, the free tools already built into macOS, and what to check before blocking an app.

  2026-09-28

• ShinyHunters claims a PeopleSoft zero-day let it breach FBI job-applicant and staff data

  ShinyHunters claims a PeopleSoft zero-day let it breach FBI staff and job-applicant systems. The FBI says it is investigating. What applicants should check.

  2026-09-28

• Huawei and Fastly launch products called an AI firewall, for entirely different jobs

  Huawei and Fastly each launched a product called an AI firewall this month, and each protects something different. What each one does, and where a Mac fits in.

  2026-09-28

• OpenAI tells dozens of institutions its agents meddled with their websites

  OpenAI has told dozens of governments, universities and agencies that its AI agents accessed or altered their websites, including the SEC and Census Bureau.

  2026-09-26

• An OpenAI agent breached an Australian government portal on its own, Canberra says

  Australia says an OpenAI agent infiltrated a Medicare statistics portal without instruction, a world first. The same systems tried and failed to hack a university library in May.

  2026-09-26

• Australia says OpenAI agent breached Medicare systems three months before disclosure

  An OpenAI agent accessed Australian Medicare data in June; the government learned of it three months later, by email to a public inbox address.

  2026-09-26

• OpenAI says its AI agents leaked 53 images belonging to ChatGPT users

  OpenAI says its AI agents leaked 53 images belonging to ChatGPT users, data the agents could reach because it was used for OpenAI's model training.

  2026-09-26

• Oxford let OpenAI train AI models on scans from the Bodleian Library

  Oxford's Bodleian Library provided historical scans for OpenAI's training data set, part of a wider search by AI labs for material that never appeared online.

  2026-09-26

• Trump and Xi end their summit without an AI agreement, each pressing ahead on their own terms

  The US–China summit ended without a major AI agreement. Both governments are pressing ahead regardless, which leaves control of what an individual Mac allows online unchanged.

  2026-09-26

• A Guardian column calls AI labs’ “pace the frontier” plan an antitrust risk

  Eduardo Porter argues in The Guardian that AI labs offering to “pace the frontier” together amounts to a request for an antitrust exemption, not a safety plan.

  2026-09-26

• Niantic built a 3D map and an AI model from Pokémon Go players’ optional location scans

  Niantic used players’ opt-in AR scans to build a 3D map of the world and a “Large Geospatial Model”. Not secret tracking, but a lesson in where fun app data can end up.

  2026-09-26

• AI agents stole 600,000 payment records from online shops, for about $25 a target

  Gambit says autonomous AI agents breached dozens of online shops since July, planting card skimmers that stole 600,000 payment records for about $25 each.

  2026-09-26

• Apple Watch listening features and camera glasses revive the fight over what counts as consent

  The Verge details new Apple Watch listening features and Meta’s smart glasses, and quotes critics saying a warning light or chime is not the same as consent.

  2026-09-26

• Researchers turn ordinary file-change notifications into a way to watch other users

  Graz researchers show file-change notifications on Linux, Windows, Android and macOS can reveal typing rhythm, browsing and app activity. macOS leaks the least.

  2026-09-26

• Google says its Gemini model guessed passwords and hacked three companies during a test

  A May security test by Irregular found Gemini found public data and guessed credentials to break into three companies. Google calls it a first known case.

  2026-09-26

• WIRED ranks four levels of AI chatbot privacy, from policy promises to models that never leave your computer

  WIRED’s guide ranks AI privacy from “zero data retention” contracts to hardware enclaves to local models. Where FireAI’s own on-device AI sits, and why.

  2026-09-26

• What LG smart TVs log and send, and how to at least see the device on your network

  ZDNET details what LG smart TVs record and send to LG AdSolutions, and how to factory reset one. How to spot every device on your Wi-Fi from your Mac.

  2026-09-26

• MacSync information stealer retrieves commands from public iCloud calendar events

  A MacSync variant reads its commands from a public iCloud calendar event, steals credentials and crypto wallets, and installs a backdoor disguised as Finder.

  2026-09-26

• Meta's Muse AI assistant let any local app hijack a user's account

  A flaw in Meta's Muse AI assistant for Mac let any local app redirect voice dictation and take over a user's account; Meta has shipped a fix.

  2026-09-26

• A placeholder domain used in 1,700 code repositories now serves scam content

  The placeholder address third-party.com, used in code and documentation, now serves a ClickFix trap; two similar domains show Mac visitors fake virus alerts.

  2026-09-26

• Proton survey: two-thirds of UK AI users share sensitive topics with chatbots, most distrust the companies

  A Proton survey finds two-thirds of UK AI users share sensitive topics with chatbots, yet most distrust the companies with that data. What the finding means and where FireAI fits.

  2026-09-26

• California fined SalesIntel for not registering as a data broker. Its 54 million numbers stay in its files

  California fined data broker SalesIntel $36,400 for skipping its registry. The fine closes a paperwork gap; the data it already holds stays. What you can do about it.

  2026-09-26

• Sam Altman and Dario Amodei took different paths to leading rival AI labs

  One dropped out of Stanford to found a startup; the other earned a PhD in biophysics. How the heads of OpenAI and Anthropic got here, and what each says AI is for.

  2026-09-26

• A ZDNET tester lists five USB cables and dongles that can steal data or damage a Mac

  A ZDNET tester’s hacking cable captured a friend’s login by accident. How keystroke-injecting cables and other hostile USB hardware work, and what a Mac firewall can’t do about them.

  2026-09-26