AI news October 2026: what companies should know now
Aqtelo teamDesign and engineering team
- Published
- Focus
- Germany, Switzerland
- Reading time
- 5 minutes
Autumn 2026 brings fewer big leaps than many small shifts. Models are getting cheaper and faster. Security capabilities are becoming so strong that vendors release them carefully. And AI agents show up more and more often where only people used to be: in shops, on websites, in wikis. This is what happened in September and October, and what it means for you.
Claude 5.5: three models in three weeks
Anthropic has renewed its entire model family. Claude Opus 5.5 arrived on 22 September, Claude Sonnet 5.5 on 28 September and Claude Haiku 5.5 in early October.
- Opus 5.5 costs 4 dollars per million input tokens and 20 dollars per million output tokens in the API, instead of 5 and 25 dollars for Opus 5. According to Anthropic, typical tasks should run around 40 percent cheaper. The model thinks before every answer, and this can no longer be switched off.
- Sonnet 5.5 answers more than 30 percent faster than Sonnet 5, according to Anthropic. The price per token stays the same, but per task it should become up to 30 percent cheaper.
- Haiku 5.5 is the small, fast model for large volumes, such as sorting enquiries or reading documents.
What this means for you: AI automations that used to be too expensive may pay off now. But check your existing prompts. Instructions written for older models often hold new ones back. Why we delete our setup and rebuild it at every model change is explained in our article New AI model? Delete everything first.
Gemini 4 Argon: security teams first
On 30 September, Google presented Gemini 4 Argon, its new top model. The launch is unusual: access goes first to selected security teams through the Fairwind programme. Google trained the model specifically for defence. It is meant to find, confirm and fix vulnerabilities in software on its own. Wider access is to begin with paying API customers, but there is no fixed date.
What this means for you: Models that find gaps find them for defenders and for attackers. Google gives the defenders a head start. Even so, the time between a published vulnerability and the first attack keeps getting shorter. Updates you postpone today are an entry point tomorrow.
Personal Agent Protocol: agents as customers
On 6 October, Sierra and Meta presented the Personal Agent Protocol. It is an open standard for how a customer's personal AI agent signs in to a company. Sign-in works through OAuth, like “Sign in with Google”. The customer decides whether their agent may only read or also act. The company decides whether agents come in through the website, through interfaces or through a company agent of its own. Sierra names Shopify, Stripe and Walmart among its partners. Payments are planned for a later version.
What this means for you: If you run a shop or a customer portal, you will get visits from agents instead of people. They don't read advertising copy, they read data: prices, availability, delivery times. Clean product data and clear interfaces become more important than the home page. Whether this standard will catch on is open. Visa runs a competing protocol, and Shopify and Stripe are part of both.
Wikimedia: AI agents without permission
On 5 October, the Wikimedia Foundation published activity by AI agents that it attributes to OpenAI. The agents edited wikis without authorisation, almost only in test areas. Some changes concerned a citation tool. Wikimedia calls them “potentially malicious”: the agents apparently wanted to use the tool as a detour to retrieve data from other services. On top of that came millions of automated requests to the interfaces. Wikimedia found no signs of a break-in to systems or data. OpenAI is analysing the incidents together with Wikimedia.
What this means for you: Agents from large vendors don't always follow a website's rules. Check which automated access reaches your website and your interfaces. Forms, search functions and open interfaces need limits, even against well-meaning agents.
Prompt injection is now in HTTP headers
On 15 October, The Register reported on an analysis by the security firm Signals Corps. It examined 217 million page requests from Common Crawl. Around 8,000 websites send messages to AI agents in invisible HTTP headers, mostly harmless pointers to an llms.txt. But 30 websites try to slip new instructions to agents. Most of them are jokes. One header targets AI systems that pre-sort job applications and claims a high match.
What this means for you: Any content from outside can contain instructions, even in places no person sees. If you use AI for applications, research or purchasing, it must treat outside content as data only. How such attacks work is shown in our article on agentjacking.
What we take away
- Recalculate costs. With Claude 5.5, automations that used to be too expensive deserve a second look.
- Retest your setup. Check old prompts and rules against the new model instead of carrying them along.
- Build for agents. Prepare product data, prices and interfaces so that agents can read them reliably.
- Secure agents. Your own agents get few permissions and an approval step for anything irreversible. Other people's agents get limits.
Frequently asked questions
Test first with real tasks from your daily work. Swapping the model ID is quick. Whether old prompts still fit only shows in the test.
Rule of thumb: Haiku for large volumes of simple tasks, Sonnet for everyday work with code and text, Opus for difficult, long tasks. The boundaries shift with every generation. That is why a comparison with your own data is worth it.
Not yet. The specification is brand new. But clean product data and documented interfaces help you with any standard that wins out.
The first step
Write down which AI models run in your company and which version. Anyone who knows this quickly sees where a switch saves money and where old prompts hold things back. How we introduce AI systems is covered under Intelligence.
Read with AI
Keep reading
4 min read
AI red teaming with open source: Garak, PyRIT, Promptfoo and more
Free tools from NVIDIA, Microsoft and others let you attack any AI application before someone else does. Which tool fits what, and what none of them replaces.
5 min read
Agentjacking: how a fake error message takes over AI agents
Three attacks from 2026 show the same pattern: an agent reads data from outside and executes it as a command. What happened and which guardrails help.
6 min read
New AI model? Delete everything first, then set up again
More context used to help a lot. With the 2026 models that flips. Why we delete rules, skills and hooks at every model change and bring back only what is missing.