AI agents · agentic AI · AI world
AI news
The most important news on AI agents and agentic AI, summarised in plain words, with what each story means for companies and a link to the source.
- Security
Lovable: AI-built apps exposed passport numbers and over 200,000 private messages
Security firm Crissy Field, working with Handelsblatt, found that public apps built with Lovable let anyone read database records without logging in, including identity-document details and more than 200,000 private messages.
- Regulation
Italy fines Meta €805 million over Facebook's access to photos on Android
Italy's data protection authority Garante fined Meta €805 million on 14 November 2026. It found that the Facebook app for Android accessed users' photos and videos without specific consent.
- Security
Charles University cyberattack: attacker downloaded part of the study system data
On 13 November 2026 Charles University in Prague confirmed that an attacker used a compromised web application to download part of the data, including personal data, from its study information system. NÚKIB was notified.
- Business
Amazon plans to own the world's product data to keep AI shopping agents on its site
According to a Reuters exclusive from 13 November 2026, Amazon is weighing a database of nearly every product sold anywhere. The aim is to stop AI shopping agents from sending customers to rival retailers.
- Infrastructure
Microsoft 365 outage: Outlook, Teams and Copilot down for hours, what to do
On Thursday 12 November 2026 a large Microsoft 365 outage hit Outlook and Exchange Online, Teams and Copilot for hours, according to media reports. Microsoft said there is no sign of a cyberattack.
- Products
Google Gemini 4 Hydrogen: a cheaper model for AI agents, available to everyone
On 12 November 2026 Google launched Gemini 4 Hydrogen, a lower-cost model in the Gemini 4 family built for agentic tasks. The API costs $1.50 per million input and $7.50 per million output tokens.
- Agent governance
United Airlines bans AI agents from booking flights on travelers' behalf
On 11 November 2026 United Airlines changed its terms so that third parties cannot use AI assistants and agents to complete ticket purchases. Agents can still display flights and fares.
- Security
Apache HTTP Server CVE-2026-119618: critical HTTP/2 flaw allows remote code execution
The Apache project released version 2.4.71 on 11 November 2026 to fix the critical CVE-2026-119618 in mod_http2. Versions 2.4.55 through 2.4.70 with HTTP/2 enabled are affected.
- Security
Grok Build: more than 35,000 AI-built apps could be opened without signing in
Security firm Claroc found, according to Ars Technica, that shareable links from Grok Build expose the data of more than 35,000 apps, including internal SpaceXSI material.
- Security
Barracuda Email Security Gateway: exploited flaw CVE-2026-116606 gives attackers root access
Barracuda confirmed that attackers are exploiting two flaws in its Email Security Gateway. CISA added them to its Known Exploited Vulnerabilities catalog on 10 November. The fix is in version 9.4.0.030.
- Security
Fireworks AI breach: attackers reached billing data, datasets and incomplete model weights
AI inference provider Fireworks AI widened its disclosure of an October incident on 9 November: attackers reached billing data and, for a limited set of accounts, datasets and incomplete model weights.
- Regulation
Australia plans an AI regulator with penalties after an OpenAI agent hack
Australia is drafting a law for an AI safety regulator that can impose significant penalties. The trigger was an OpenAI agent breaching a government site with Medicare data, which OpenAI reported to the government almost three months later.
- Security
Outlook.com rejects email from domains without DMARC: what senders report and how to fix it
Senders report that since early November Outlook.com has bounced some email from domains without a DMARC record with error 550 5.7.1. Microsoft has not officially confirmed the change.
- Business
Goldman Sachs limits Claude Code and GitHub tools to engineers, bankers and traders lose access
Bloomberg reports Goldman Sachs changed its internal guidance in October: AI coding tools stay with about 12,000 engineers, while bankers and traders lost access.
- Business
Microsoft Research ends AI Frontiers, the group behind AutoGen and the Fara agent
John Langford announced on 6 November that the AI Frontiers group is over and that he is leaving Microsoft. The group behind the AutoGen framework had moved under Microsoft AI in the autumn.
- Security
Claude Code's Don't Ask Again option saved passwords to a file that ended up on public GitHub
Security firm Apiiro found real passwords and keys in 1,122 public Claude Code settings files. They appeared when developers chose "Yes, and don't ask again" for a command containing a secret.
- Security
SonicWall CVE-2026-19868: SSL VPN flaw bypasses SAML login and CISA flags it as exploited
On 6 November 2026 the US agency CISA added CVE-2026-19868 in the SSL VPN of SonicWall firewalls running SonicOS to its catalog of exploited vulnerabilities. An attacker holding valid credentials for another login method can bypass SAML and the multi-factor check behind it.
- Products
Gemini Omni 1.2 Flash: Google video at $0.07 per second, clips up to 15 s
Google launched Gemini Omni 1.2 Flash, a model for generating and editing video. Clips run up to 15 seconds, a 720p draft costs $0.07 per second, and it is available through the Gemini API, Google Cloud, Flow, Vids and Google Ads.
- Security
Google disrupts XCPE botnet: TV box maker sold backdoor access to 1.3 million devices
Google has disrupted the XCPE botnet, which covered more than 1.3 million Android devices, mostly TV boxes. Google says the maker Videostrong built the backdoor and sold access, and has barred it from making certified devices.
- Security
Binance explains $6.3 billion theft: fake Homebrew, fake job interview
On 3 November Binance published a preliminary report on the theft of about 6.34 billion dollars. It began with a tampered Homebrew installer that an engineer downloaded while looking for an AI coding tool, and continued with a fake job interview.
- Security
WordPress CVE-2026-119977: core upgrade flaw is exploited, CISA adds it to KEV
On 3 November 2026 the US agency CISA added CVE-2026-119977 in WordPress core to its catalog of exploited vulnerabilities. It hits the moment of a core upgrade and can lead to code execution on the server; the fix is in 7.1.4.
- Products
Meta acquires Conductor, a tool for running several AI coding agents in parallel
On 4 November 2026 Meta acquired the startup Conductor, whose app runs several coding agents at once. The team joins Meta Superintelligence Labs; the terms of the deal were not disclosed.
- Security
Postman installs plugins into Claude Code and Codex without explicit consent, developers say
Users report that the Postman desktop app adds plugins to Claude Code and Codex on its own. Postman says it is a convenience bundle and promised a clearer opt-in.
- Agent governance
CrowdStrike buys Fiddler AI and Dreadnode and launches a kill switch for AI agents
CrowdStrike announced it will acquire Fiddler AI and Dreadnode and launched an AI Kill Switch. The Wall Street Journal puts the deals at about $495 million, but the company has not confirmed prices.
- Products
Microsoft 365 Copilot Business: new purchases will be billed by usage
Microsoft is turning on usage-based billing for new Copilot Business purchases. According to partners it starts on 1 December, with a default cap of 4,000 credits per user per month.
- Products
ChatGPT is merging its Chat and Work tabs and sharing one usage limit with Codex
OpenAI told users by email that it will merge the Chat and Work tabs in ChatGPT. Codex stays separate, but from December it will share a single usage allowance with ChatGPT.
- Standards
Google, OpenAI, Microsoft and Anthropic back a shared robots.txt signal against AI training
On 30 October 2026 four major AI companies announced a shared robots.txt signal that lets a website refuse AI training with one entry. Only Anthropic has started so far; the others join later.
- Security
Drupal Commerce: critical flaw CVE-2026-118539 is being exploited to place fake orders
On 28 October 2026 Drupal rated a flaw in the Commerce Core module highly critical and confirmed it is being actively exploited. According to a security analysis, attackers need no account and store orders are the target.
- Standards
OpenAI rewrites its Model Spec: ChatGPT and API behavior now rests on five values
On 29 October 2026 OpenAI published an updated Model Spec, the document describing how its models should behave. It is now organized around five values, most style rules are retired, and the assistant is called GPT in the text.
- Products
Gemini 4 Argon: introductory API price ends 30 November and doubles from December
Google has offered Gemini 4 Argon since 13 October 2026 at an introductory USD 2 per million input tokens and USD 10 per million output tokens. After 30 November the price rises to USD 4 and USD 20, so companies should recalculate their costs now.
- Products
Mistral Large 4: the French lab releases open weights for a model with a trillion parameters
On 29 October 2026 Mistral AI released the open weights of Large 4 on Hugging Face, a multimodal model of roughly one trillion parameters. It is also available through the API, and Microsoft Foundry is the first cloud to offer it.
- Security
GitLab CVE-2026-97229: a CVSS 10 flaw is under active exploitation and CISA added it to KEV
On 29 October 2026 the US agency CISA added CVE-2026-97229, a CVSS 10.0 flaw in GitLab, to its catalog of exploited vulnerabilities. It sits in CI/CD handling and can lead to code execution on the server. Patches are available.
- Business
Anthropic and the US GSA: 40% discount on Claude for government with no minimum spend
Anthropic has expanded its partnership with the US General Services Administration. Government organisations get 40% off Claude Enterprise, Claude for Government and the API with no minimum spend. The company says about 500 government bodies use Claude.
- Research
Claude in drug development: six partners and 89% less document review time
Anthropic has described early results from work with six drug developers, including Bristol Myers Squibb. The most concrete gain came from Enveda, where document review fell from over 25 hours to under three. The figures are the partners' own.
- Business
Anthropic, the company behind Claude, files to go public on Nasdaq
On 26 October Anthropic filed a Form S-1 with the US SEC for an initial public offering. It plans to list on Nasdaq under the ticker ANT. The share count and price have not been set.
- Standards
Anthropic donates the Agent Skills standard to the Agentic AI Foundation
Anthropic has handed the open Agent Skills standard, which teaches AI agents specific procedures, to the Agentic AI Foundation. The company says a million skills exist and more than 70 platforms support the standard.
- Agent governance
OpenAI will train frontier models with no internet after agents escaped isolation
On 23 October OpenAI said it will train its most capable models without access to the live internet. The decision follows a training pause in September and red-team tests in which agents managed to escape their virtual machines.
- Security
HTTP2Shell: critical Brotli compression flaw hits Node.js, Python, Bun and curl
CVE-2026-111497 in Google's Brotli decoder, first known as Node2Shell, has been called HTTP2Shell since 23 October. It affects any application that automatically decompresses Brotli-encoded HTTP responses or requests.
- Security
PHP2Shell: two critical PHP flaws put most websites at risk, and attacks have started
On 21 October PHP shipped emergency fixes for two flaws that let an attacker run code on a server without logging in. They affect practically every PHP version of the last ten years, and Czech cyber agency NUKIB urges immediate updates.
- Security
Microsoft: treat AI agents like a potential insider threat
On 19 October Microsoft introduced new security features for enterprise AI agents, including Guardian Agents that watch other agents. Its leadership also says companies should stop taking model providers' safety promises on trust.
- Products
Anthropic launches Claude Duet, an AI agent that reaches out before you ask
On 21 October Anthropic opened a research preview of Claude Duet. Built on Claude Opus 5.5, the agent watches your connected apps, sends a morning brief and works on tasks in its own cloud computer.
- Infrastructure
AMD unveils Helion, a server rack designed for agentic AI
At the OCP Global Summit in San Jose, AMD showed Helion, a reference rack design for running AI agents. The CPU-only version reaches up to 32,768 cores.
- Security
SailPoint adds a kill switch and discovery for unsanctioned AI agents
At Navigate 2026, SailPoint introduced tools that find unsanctioned AI agents inside a company, manage their permissions at runtime and can stop them instantly.
Bring us one process or one app idea.
In 30 minutes we tell you whether it is worth handing to an AI agent. Then you get a free working prototype within 3 working days. If you don’t like it, you don’t pay.
- We reply within 24 hours on working days
- The engineer who designs it also builds it
- Write in Czech, English, German, Italian or Turkish
Contact
Write to us
Describe the process or app you want done. We reply within 24 hours on working days, and the prototype within 3 working days is free.