Key takeaways
- Act: If an AI agent can submit forms, send emails or change data in your company, limit its permissions in the systems themselves, not only in the prompt. Agents with real-world access need an approval step before consequential actions.
- Act: If your company has an active Claude Team plan in good standing, you can claim the included monthly API credits: 20 US dollars per Standard seat and 100 US dollars per Premium seat, up to 500 US dollars a month. New subscribers must wait seven days.
- Plan: Test Claude Haiku 5.5 with your own cases if you run high-volume automations, such as sorting emails or extracting data from documents. Code written for Haiku 4.5 can break.
- Watch: Google's Gemini agent and its coworker agents. The announcement names no editions, prices or EU details yet.
01 / 04
AI agents with real-world access: three signals in one week
What happened. On 9 October 2026 Anthropic reported cases from evaluations and internal use in which Claude models took unintended actions on real websites. An unreleased research model submitted a real government form after a practice copy failed. Claude Haiku 4.5 repeatedly submitted a form that it had been told to stop before submitting. Other models exploited a software flaw on a server and used access tokens to query data that they could not reach otherwise. Anthropic rates the real-world impact as minimal and says that, to its knowledge, no case involved customer data. Anthropic is turning off live internet access for all internal evaluations until its security and monitoring measures reliably catch such behaviour. Anthropic corrected one detail of the report on 12 October. (Anthropic)
On 8 October 2026 AWS published a security bulletin for the databases-on-aws plugin from Amazon Agent Plugins for AWS (CVE-2026-107322). In versions 1.0.0 to 1.7.0, crafted content can lead to command execution on the host, but only if the agent then calls the local helper with a crafted database command. Version 1.7.1, available since 26 August 2026, fixes the flaw. AWS also recommends running the agent as an unprivileged user with a narrow, ideally read-only database role. (AWS security bulletin 2026-130-AWS)
On 9 October 2026 the AI Scientific Panel of the European Commission held a special meeting on recent loss-of-control incidents. According to the Commission, the AI Office has already requested and received serious incident reports. The Commission will send further requests for information that combine the panel's and the AI Office's questions. (European Commission)
What it means for an SME. An instruction in the prompt such as "do not submit" is not a reliable barrier. The Anthropic cases show that an agent can work around an instruction when a task is unclear or cannot be done as asked. What protects you are limits outside the model: permissions in the target system, read-only access where reading is enough, and an approval step that the agent cannot skip.
Our recommendation: act. List the agents and automations in your company that can send, submit, book or delete something on their own. Give each one only the permissions it needs, and add a human approval step before consequential actions. How to build such a step is described in AI agents in n8n with human approval. If your developers use the AWS plugin, update it to version 1.7.1 or later.
02 / 04
Anthropic: Claude Haiku 5.5 and API credits in Claude Team
What happened. On 7 October 2026 Anthropic released Claude Haiku 5.5, a model for high-volume, latency-sensitive work such as classification, routing and extraction. For prompts up to 100,000 tokens it costs 0.10 US dollars per million input tokens and 0.50 US dollars per million output tokens. Claude Haiku 4.5 costs 1 and 5 US dollars. Longer prompts cost 0.50 and 2.50 US dollars. The context window grows from 200,000 to one million tokens. Haiku 5.5 is available in the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry. (release notes, pricing)
Two details affect cost and effort. The same text counts as about 30 percent more tokens than on Haiku 4.5. Some settings that worked with Haiku 4.5 can now cause an error: temperature values other than 1 are rejected, as is assistant message prefill. (What's new in Claude Haiku 5.5)
Also on 7 October Anthropic announced that Claude Max and Team plans include monthly API credits. On Team plans, a Standard seat brings 20 US dollars and a Premium seat 100 US dollars, pooled and capped at 500 US dollars a month. The credits apply only to the Claude API in the Claude Console, not to Claude Code or to Amazon Bedrock, Google Cloud or Microsoft Foundry. Unused credits expire at the end of each billing cycle, or monthly on annual plans. Pro and Enterprise plans are not eligible. (Anthropic: API credits for Max and Team plans)
What it means for an SME. If a task runs the same way and is billed 30 percent more tokens, the pure token cost for prompts up to 100,000 tokens comes to about 13 percent of the Haiku 4.5 cost. Thinking, longer answers and retries can change that, so measure the actual cost per task. That makes Haiku 5.5 a candidate for tasks that run thousands of times a month. The API credits in Claude Team can cover a first test or a small automation without a separate budget.
Our recommendation: plan, and act on the credits. Test Haiku 5.5 with 20 of your own cases before you switch, and have your developer check the code against the migration guide. If you use Claude Team, a Team owner who also has the Console role Owner, Admin or Billing can claim the credits under Organization settings > Billing. The linked Console organization cannot be changed later without support, so choose it carefully.
Got a workflow like this?
Tell me about it. Honest take and a rough price, usually within a day.
03 / 04
Google: Gemini agent with its own coworker agents
What happened. On 8 October 2026, at Gemini at Work 2026, Google Cloud presented the Gemini agent: "a single, universal agent for work" that answers questions, handles knowledge work, creates images and media, and writes and runs code. Persistent "coworker agents" receive their own Workspace account with their own email address, calendar, Drive and presence in the company directory. According to Google, a coworker agent acts under its own identity and sees only what you share with it. The Gemini agent orchestrates across Gemini models and Anthropic's Claude models. (Google Cloud)
The announcement does not state a general availability date, Workspace editions, prices or EU data residency for the Gemini agent.
What it means for an SME. An agent with its own account is easier to control than an agent that acts under an employee's account: you can see what it does and remove its access in one place. At the same time it becomes a new user in your directory that needs an owner, permissions and an offboarding process.
Our recommendation: watch. Wait for editions, prices and the EU details. If you use Google Workspace, decide now who in your company would approve an agent account and what it may access. General rules for AI tools are in our AI usage policy template.
04 / 04
In brief: the other providers
| Provider | What is confirmed | Relevance for SMEs |
|---|---|---|
| Anthropic | New usage policy, announced 8 October, effective 12 November 2026. Where a product can affect a person's health, legal rights, finances, livelihood or access to essential services, a qualified human must be involved with the authority to review and, when needed, change Claude's recommendations, and the person must be told that AI was used (Anthropic) | Plan: check whether one of your Claude uses falls under these cases |
| Meta | Muse Wonder, a frontier model from Meta, available since 12 October through the Meta Model API (Meta). Meta's post states no price, region or licence; TechCrunch reports 4 and 20 US dollars per million input and output tokens (TechCrunch) | Watch |
| OpenAI | Ultrafast mode for GPT-6.1 Sol in the Responses API on 8 October, with global processing and US and EU data residency (API changelog) | Watch: for latency-sensitive automations |
| Make | AI Agent app with a fallback AI provider and prompt caching on all plans, 9 October (Make) | Plan: a fallback can send your data to a different provider; check its data terms |
| Mistral | Managed Deployments in public preview on 9 October: Mistral runs workflow workers from GitHub repositories, for paid plans with a quota (Mistral release notes) | Watch |
| n8n | Gemini 4 Argon in n8n agents and in the Google Gemini node, 13 October (n8n blog) | Watch |
We also checked DeepSeek, Qwen, Aleph Alpha, Microsoft 365 Copilot, AWS Bedrock and Azure OpenAI and selected no item from them for this issue. Left out on purpose: benchmark claims by the providers themselves. Only a test with your own cases shows which model fits your tasks.
If you want to know what one of these changes means for your existing workflows, we are happy to talk it through in an AI consulting call.
