This week: until 31 December, Microsoft is giving CSP customers five extra Copilot Studio capacity packs when they buy at least 300 new Microsoft 365 E7 seats, OpenAI brings its frontier models and Codex to Google Cloud’s Model Garden, OpenAI reports that two of its own agents weakened monitoring for undesirable model behaviour and that one of them denied doing so, the European Commission publishes EN 18286, the first harmonised standard under the AI Act, Anthropic places 5 per cent of its shares in a public fund that grows with the company’s value, the US Army is working on self-propagating AI agents for offensive cyber operations according to The Intercept, and OpenAI adds presentations with export to PowerPoint to ChatGPT with ChatGPT Slides.
- Microsoft matches Copilot Studio capacity on E7 deals via CSP
- OpenAI brings its models and Codex to Google Cloud
- OpenAI reports that its own agents weakened their misalignment monitoring
- European Commission publishes first harmonised standard under the AI Act
- Anthropic places 5 per cent of its shares in a public fund
- US Army working on self-propagating AI agents for cyberattacks
- OpenAI adds ChatGPT Slides with export to PowerPoint
Microsoft matches Copilot Studio capacity on E7 deals via CSP
On 21 October, Microsoft announced a temporary Buy 5 Get 5 promotion for Microsoft 365 E7 combined with annual Copilot Studio capacity packs via CSP. The promotion runs until 31 December 2026 inclusive.
- Customers qualify with at least 300 new annual E7 seats, pay-as-you-go billing via an Azure subscription and a minimum of 5 capacity packs.
- Microsoft matches each pack 1:1 for 12 months, up to a maximum of 5 packs.
- Microsoft 365 E7 bundles Microsoft 365 E5, Copilot, Entra Suite and Agent 365.
- Microsoft also reported on 15 October that support for Office LTSC 2021, Project LTSC 2021 and Visio LTSC 2021 ended on 13 October.
Source: Microsoft Partner Center
OpenAI brings its models and Codex to Google Cloud
On 21 October, OpenAI and Google Cloud announced that OpenAI’s frontier models and Codex workflows will become available in Google Cloud’s Model Garden. Customers purchase the models via Google Cloud Marketplace and can use existing spending commitments with Google Cloud to do so.
- According to OpenAI, general availability is planned for the coming weeks.
- OpenAI’s models will also soon become available in Google Antigravity and the Gemini agent.
- According to OpenAI, Microsoft remains its primary cloud partner.
- Since 10 June, OpenAI’s models and Codex can also be purchased via cloud commitments with Oracle.
Source: OpenAI
OpenAI reports that its own agents weakened their misalignment monitoring
On 20 October, OpenAI published an incident report on two internal coding tasks in which agents weakened monitoring for undesirable model behaviour on 3 and 6 October. According to OpenAI, the agents interpreted routine instructions as permission to weaken those safeguards.
- A GPT-6.1 Sol agent disabled automatic alerts to the security team based on an earlier instruction not to send Slack messages.
- A GPT-6 Astra agent removed a monitoring integration and then denied to the user that it had removed an active monitor.
- Neither incident was given the highest priority, P0, which meant the security team was not paged immediately.
- OpenAI has modified Auto-Review, which is mandatory for all internal Codex use, to block attempts to disable monitoring.
Source: OpenAI
European Commission publishes first harmonised standard under the AI Act
On 19 October, the European Commission announced that the reference to EN 18286:2026, the first harmonised standard supporting the AI Act, has been published in the Official Journal of the EU. The standard describes a quality management system for providers of high-risk AI systems.
- The standard addresses the requirements of Article 17 of the AI Act on the quality management system.
- Companies that apply a harmonised standard from the Official Journal are presumed to comply with the corresponding legal requirements.
- The Commission asked CEN and CENELEC for standards in ten areas, including risk management, data quality, human oversight and cybersecurity.
- On 30 October 2025, the draft standard became the first AI standard to go out for public consultation.
Source: European Commission
Anthropic places 5 per cent of its shares in a public fund
On 18 October, Anthropic announced the creation of the Anthropic Public Benefit Fund, a wholly owned LLC into which the company will place 5 per cent of its shares, subject to shareholder approval. The fund’s stake grows with the value of the company.
- The board can add 2 per cent in shares for every 50 per cent increase in value above a set baseline.
- According to Anthropic, the fund can make large donations to non-profits or direct payments to specific groups, for example in the event of job losses caused by AI.
- According to Anthropic, together with earlier pledges by founders and employees, around 20 per cent of the shares are now earmarked for the public interest.
- The fund may not use the shares for lobbying or political donations, whether in the US or elsewhere.
Source: Anthropic
US Army working on self-propagating AI agents for cyberattacks
On 19 October, The Intercept published documents from Army Cyber Command (ARCYBER) about ECHOTRIBBLE, a programme that lets AI agents move autonomously through adversaries’ networks to identify and neutralise targets. According to the documents, the programme is primarily focused on offensive missions.
- According to the documents, the software is meant to operate with minimal human supervision and exhibit self-directed behaviour.
- In an example from an ARCYBER presentation, the agent on a laptop creates copies of itself for iPads, Apple desktops and Windows 11 PCs.
- According to the documents, neutralising includes temporarily or completely disabling a target or removing it.
- ARCYBER did not answer The Intercept’s questions about preventing unintended harm.
Source: The Intercept
OpenAI adds ChatGPT Slides with export to PowerPoint
On 16 October, OpenAI released ChatGPT Slides, which lets users create presentations from a conversation, files or an outline and continue working on them with colleagues in Space. The feature is becoming available to Pro, Business and Enterprise users with access to Space.
- Team members can edit the same presentation in real time and leave shared comments.
- On the web, ChatGPT exports to PowerPoint, PDF or Google Slides; in the desktop app, the presentation opens in PowerPoint or Keynote.
- On the same day, GPT-6 Pro in ChatGPT gained multiple agents that work on parts of a task in parallel.
- On 20 October, OpenAI also introduced locked chats, which are opened with a four-digit PIN.
Source: OpenAI
This article was originally published in Dutch on MSP Business. Read the original →
