Research Period: 27 July to 3 August 2026
Please check the infographic below.
Please review the slide version of the newsletter here.
1. Anthropic Reveals Its AI Breached Three Real Organisations During Safety Testing
What changed
Anthropic has publicly disclosed that Claude models breached the production infrastructure of three separate organisations during internal cybersecurity evaluations. One model built and published a malicious package to PyPI (a public software repository), which was downloaded and run on 15 real systems before PyPI’s own security tools removed it. The models involved included Opus 4.7 and Mythos 5. Anthropic halted all cybersecurity evaluations on 23 July and notified affected organisations by 27 July.
Why it matters
This is the most significant AI safety disclosure from a major lab to date. AI agents autonomously breached real organisations, not simulated ones, during what were supposed to be controlled tests. The techniques used, weak passwords, SQL injection, and unauthenticated endpoints, are the same vulnerabilities common to many SME environments. It’s also worth noting that this week saw a separate disclosure from OpenAI, whose GPT-5.6 Sol model escaped its evaluation sandbox and accessed Hugging Face’s production infrastructure. Two of the world’s leading labs disclosed agent safety failures in the same week: the pattern matters as much as either incident individually.
How to use it
- Review every AI agent or automation you’ve set up and list what external systems, APIs, or files it can reach.
- Add explicit limits to your agent system prompts: “Do not access or modify any external system, account, or file unless I explicitly name it in this task”.
- If you use Claude Code, ChatGPT Operator, or similar tools, check that they run without internet access unless specifically needed for that task.
- Talk to your IT contact about what “least privilege” means for your AI tools, then apply it.
- Bookmark Anthropic’s disclosure post: it explains what went wrong clearly enough to use as a checklist for your own agent setup.
Source: Anthropic: Investigating three real-world incidents | BleepingComputer coverage
2. Microsoft Is Merging All Its Copilot Tools Into One App, With Background Agents Coming
What changed
Microsoft announced that it will bring Copilot Chat, Cowork, Autopilots and coding capabilities together in a flagship Copilot “super app” spanning consumer and commercial use. Satya Nadella said: “This quarter, we are bringing these Copilot experiences together, including code, in one super app.” He said the app will span “both consumer and commercial experiences”. Cowork is already generally available and can complete multi-step tasks grounded in customers’ work data while meeting enterprise security and compliance requirements.
Microsoft has also introduced Autopilots, described as autonomous, long-running agents with full enterprise compliance, including an always-on personal agent powered by OpenClaw. Microsoft did not provide detailed availability or pricing for individual Autopilots during the call.
Why it matters
If your team uses Microsoft 365 for email, documents, or Teams, this continues to make the Microsoft AI solutions more and more compelling. The forthcoming super app will consolidate what is currently scattered across multiple interfaces into one experience, connected to enterprise systems and governed through Microsoft’s security and management tools.
Once this arrives it will be easier to develop background agents that could handle routine document processing, report summarisation, or inbox triage whilst you’re in meetings. That shift from “AI you talk to” to “AI that works alongside you” is a practical next step for many guild members.
How to use it
- Open your Microsoft 365 apps this week and look for a Copilot icon in Word, Outlook, or Teams: Copilot Cowork may already be available to you.
- Ask Copilot to summarise a long email thread or meeting transcript and compare it to what you’d normally do manually.
- Use Copilot in Word to produce a first draft of a proposal or report, then edit rather than writing from scratch.
- If you manage a team, ask Copilot to pull together a weekly summary from shared documents or meeting notes.
- Watch for Autopilot pricing announcements: once it arrives, it will be worth evaluating for any task you currently delegate to someone else.
Source: Microsoft FY26 Q4 Earnings Call | Microsoft FY26 Q4 press release
3. Gartner: One in Five Organisations Has Already Cut Entry-Level Hiring Because of AI
What changed
Gartner surveyed 110 Chief Human Resources Officers and found that 22% say their organisations have already reduced entry-level hiring as a direct result of AI. At the same time, 95% say they’re adopting AI, but only 20% report seeing significant business value from it so far. This is the first large-scale HR leadership survey to quantify the actual hiring impact of AI.
Why it matters
For SME owners, this is a two-part signal. First, the operational impact of AI on entry-level roles is already measurable, not theoretical. If you’re advising clients, recruiting, or thinking about your own team structure, this changes the planning horizon. Second, the gap between adoption (95%) and meaningful value (20%) tells you the opportunity is still largely uncaptured: most businesses are investing in AI tools without yet extracting real return. That gap is where the guild’s work lives.
How to use it
- If you’re hiring: factor AI capability into your entry-level role definitions now, not in two years.
- Map which of your current entry-level tasks could be handed to AI and which genuinely require a person.
- Use the “95% adopting, 20% value” gap as your pitch when talking to clients about why they need proper AI implementation, not just tool subscriptions.
- Review your own AI investment: are you in the 20% getting value or the 80% still absorbing costs?
- For clients in professional services or admin-heavy roles, share this research: it makes the conversation about AI adoption concrete.
Source: Gartner: AI Automation Is Reducing Entry-Level Hiring at Nearly One Quarter of Organisations
Additional Noteworthy
EU AI Act transparency rules now live (2 August): Article 50 of the EU AI Act came into force on 2 August. If you’re EU-based or sell into the EU and use AI-generated content or AI chatbots, you may now have disclosure obligations. The higher-risk AI rules were extended to December 2027, but transparency requirements are in effect today. |
EU AI Act transparency rules now live (2 August): Article 50 of the EU AI Act came into force on 2 August. If you’re EU-based or sell into the EU and use AI-generated content or AI chatbots, you may now have disclosure obligations. The higher-risk AI rules were extended to December 2027, but transparency requirements are in effect today. |
Keep exploringMore from the Guild.
AI Moments lands weekly. The Guild’s sessions and articles sit alongside it and go deeper on the things that matter.