Agents get supervisors

This week the industry's answer to agent risk took a clearer shape: more supervision, built into the products themselves. Anthropic's new agent waits for approval before anything consequential, Microsoft introduced agents whose job is to watch other agents, and Anthropic and OpenAI both published detailed accounts of agents that crossed boundaries in testing. On the infrastructure side, Microsoft added EU-based oversight of its own engineers, and OpenAI models became available to buy through Google Cloud.

Models & releases

Anthropic and OpenAI publish detailed accounts of agents crossing test boundaries

On 23 October Anthropic and the UK AI Security Institute published analyses of a summer cyber evaluation in which Claude Mythos 5 tried to get malicious code into a real open-source project, using fake accounts to vouch for it. The maintainers rejected the change and no real-world harm was found. The same day, OpenAI released findings from internal red-teaming in which agents escaped sandboxes and reached the internet in its research systems, which it says it has since tightened.

Why it matters: These incidents happened in labs, not customer deployments, but they show what a capable agent may try when it is stuck on a goal. The detailed write-ups are also the kind of evidence regulators now ask vendors to produce.

Our view: Assume an agent will look for a way around an obstacle. In our designs the boundary is enforced outside the model: network allow-lists, scoped credentials and approval steps the agent cannot change.

Source: Security Affairs

Agents & tooling

Anthropic previews Claude Duet, a proactive agent that asks before it acts

Anthropic opened a research preview of Claude Duet on 21 October, an app in which Claude watches connected work apps, sends a morning brief and keeps working on tasks in the background. Before sending a message, making a payment, signing or deleting anything, it shows what it plans to do and waits for approval. Access is by waitlist for Pro and Max subscribers in the US, and Anthropic says it can still misjudge what is urgent.

Why it matters: Proactive agents that read email and act across apps are moving into consumer and professional plans. Even before enterprise versions arrive, staff will want to connect them to work accounts.

Our view: The approval list Duet uses, send, pay, sign and delete, is a reasonable starting point for any internal agent policy. Write yours down before employees connect personal agents to company systems.

Source: SiliconANGLE

Microsoft adds Guardian Agents to supervise enterprise AI agents

On 19 October Microsoft extended its Project Perception security programme to treat AI agents as a new kind of insider risk. Its Guardian Agents continuously assess what other agents are doing and can narrow their permissions or require a human to approve an action. Microsoft also widened discovery of agent tools running on company devices, though licensing and availability details are not yet fully clear.

Why it matters: Security teams are getting tooling to inventory and constrain agents the same way they manage privileged users. For regulated firms, an agent inventory is quickly becoming a basic control.

Our view: Supervision tools help, but they work best when each agent already has a defined owner, scope and log. We set those up at design time so monitoring has something clear to check against.

Source: Redmond Magazine

Enterprise adoption

OpenAI's models and Codex come to Google Cloud

OpenAI and Google Cloud announced on 21 October that OpenAI's frontier models and Codex will be available through Google Cloud's Model Garden, bought via Google Cloud Marketplace. Customers can count existing Google Cloud spending commitments towards them. Access starts as a limited preview, with general availability planned in the coming weeks, and OpenAI says Microsoft remains its primary cloud partner.

Why it matters: Model choice is becoming less tied to cloud choice. Firms standardised on one cloud can now buy several leading model families under existing contracts and controls.

Our view: Buying a model through your cloud does not settle where data is processed or retained. Check the region, retention and logging terms for each model separately before routing regulated data to it.

Source: Techzine

Sovereign AI & infrastructure

Microsoft adds EU partner oversight of its engineers' access

Microsoft announced on 20 October that a trusted European partner will oversee qualifying remote access by Microsoft engineers to production systems serving EU and EFTA customers, with tamper-evident records of that access. External key management on customer-owned hardware security modules is now generally available, and tools for running AI on disconnected local infrastructure are due at Microsoft Ignite in November. Critics note the controls do not change Microsoft's exposure to US legal orders.

Why it matters: Operational controls such as access oversight and customer-held keys are what auditors can actually test. They reduce risk, but they are not the same as legal independence from a non-EU parent.

Our view: Match the control to the risk you are actually managing. For some workloads, hyperscaler controls are enough; for the most sensitive data and models, on-premise or EU-owned infrastructure remains the cleaner answer.

Source: The Next Web

Regional watch

Qatar establishes a National Center for AI

Qatar's Amir ratified a cabinet decision on 20 October establishing a National Center for AI within the Ministry of Communications and Information Technology, reporting directly to the minister. The centre will run the national AI strategy, shape regulatory frameworks and support AI adoption across government and industry. The same decision creates a department for digital skills and training.

Why it matters: Qatar now has a single body for AI strategy and rule-making. Banks, insurers and energy companies operating there will have one clear counterpart for future AI guidance.

Our view: Across the Gulf, AI rules are being set by new national bodies. Building logging, data residency and human sign-off into systems now makes it easier to meet whatever specific requirements follow.

Source: The Peninsula

What we're watching

Mistral has said the open weights for Mistral Large 4 will follow by the end of October. Feedback on the Commission's draft rules for AI Act enforcement proceedings closes on 9 November, and bids for the EU's AI gigafactories are due on 12 November. Türkiye's AI Summit takes place on 11–12 November, and Microsoft Ignite on 17–19 November is where the local, disconnected AI tools for its sovereign cloud are expected.