Skip to content
IIAIT

AI Today · 20 October 2026

Securing AI agents, the first AI Act standard and a costly API key breach

Today: Microsoft's case for treating AI agents as insiders, the first harmonised standard under the EU AI Act, a breach that exposed stored API keys, longer-running agents in Notion, and new speech models from Mistral.

  1. 01

    Microsoft: AI agents at work should be secured like insiders, not trusted by default

    On 19 October Microsoft argued that AI agents working inside organisations act as insiders, with access to data, tools and systems, and are often given more permissions than they need. Hayete Gallot, who leads Microsoft Security, wrote that frontier models cannot simply be trusted, not because they are malicious, but because training to reach results can lead them to take shortcuts. Microsoft described new controls, including permissions limited to a single task that expire when it ends, human approval for higher-impact actions, continuous risk scoring for agents and an isolated environment for running them. It says it can now discover 60 types of local agents, including Claude Code, OpenAI Codex and GitHub Copilot. No release dates were given.

    What it means for you

    If you use AI agents that can read files, send messages or change systems, give each one only the access its task needs and keep a person approving consequential actions. It is worth asking your IT team which agents are running on company devices and what they are allowed to do.

    Sources

  2. 02

    First harmonised standard under the EU AI Act covers quality management for high-risk AI

    On 19 October the European Commission announced that the reference to the first harmonised standard supporting the AI Act has been published in the Official Journal, through Implementing Decision (EU) 2026/2274. The standard, EN 18286, sets out a quality management system for AI Act purposes and is designed to help providers of high-risk AI systems meet Article 17. Applying it is voluntary; providers who do are presumed to comply with the requirements it covers, and other ways of demonstrating compliance remain possible.

    What it means for you

    For organisations that build or supply high-risk AI systems, this is the first official reference point for organising compliance work. Most people who only use AI tools are not directly affected, but it is a useful question for suppliers: how do you meet the AI Act's quality management requirements?

    Sources

  3. 03

    OpenCode breach exposes stored API keys and leads to unauthorised AI charges

    OpenCode, an AI coding tool, disclosed that attackers reached API keys that users had stored in its web Console, affecting 1.4 per cent of Console users, according to TechCrunch. Developers reported unauthorised charges on their OpenAI, Anthropic and Google accounts, in a few cases over 1,000 dollars. Anthropic and Google blocked keys they believed were exposed, OpenAI refunded credits used in the attack, and OpenCode says refunds are underway. The open-source app and command-line tool were not affected. How the attackers got in has not been explained, and the company described the incident as ongoing.

    What it means for you

    An API key works like a payment card for AI services. Store keys in as few places as possible, set spending limits and alerts with each provider, and replace any key you have entered in a third-party service that reports a breach.

    Sources

  4. 04

    Notion agents can now work on one task for more than 24 hours, with new credit-based pricing

    Notion announced on 19 October that its agents can work on a single task for more than 24 hours, up from five minutes. A new Review for me setting uses a separate model to check whether an action is safe before the agent takes it. Pricing moves to monthly credits: each Business and Enterprise seat includes 1,500 Standard and 500 Select credits. New customers are on the new system now, and existing customers move over on 2 November.

    What it means for you

    Longer-running agents can take on bigger pieces of work, but they also act for longer without you watching. Start with low-risk tasks, turn on review before actions, and if your team uses Notion, check before 2 November how the credit model will affect your usage and costs.

    Sources

  5. 05

    Mistral releases Voxtral 3 speech models, available through European endpoints

    Mistral released Voxtral 3, a family of speech models for transcription, real-time transcription and text-to-speech, together with an SDK for building voice agents. The company says transcription covers more than 50 languages, including the 24 official EU languages, and that speech generation covers more than 35 languages. Batch transcription starts at 0.0035 dollars per minute. The models are available through Mistral's European, US and global API endpoints, and self-deployment is supported.

    What it means for you

    If you transcribe meetings or calls, a provider with European hosting can make data protection questions easier to answer. Test accuracy on your own recordings and language before relying on it, and check whether recording and transcribing a conversation needs the participants' consent.

    Sources