1 November 2026 · Agents move into action

AI agents are taking on increasingly operational tasks. This morning’s selection looks at what it takes to control them: training-data experiments, cyberattacks observed in the wild, isolated cloud environments and models running locally. Four original sources, with their dates and limitations kept distinct.

Anthropic · Seemingly harmless data puts safeguards to the test

Research published by Anthropic in October shows how purpose-built training data can pass a monitor and induce unwanted behaviour in a small experimental model. The system generating that data could repeatedly test against the same safeguard, a setup favourable to the attacker. The resulting models were never deployed: the study looks for weaknesses in defences, rather than documenting a compromise of Claude used by customers.

Anthropic · training monitor tests, October 2026 ↗

Google and Mandiant · Testing tools used in intrusions

In their 30 October report, Google Threat Intelligence Group and Mandiant describe the abuse of agentic frameworks originally designed for security testing, including ARTEX. The activity was observed during incident response and can automate several intrusion stages. Familiar weaknesses remain relevant: compromised credentials and unpatched systems are still entry points. Speed and scale change. The defences highlighted include least privilege, short-lived credentials, segmentation and behavioural detection.

Google GTIG and Mandiant · agentic offensive frameworks, 30 October 2026 ↗

Google Cloud · More isolation and governance for agents

Google Cloud’s October update presents Agent Compute Engine, with microVM-isolated execution environments, and tools including Agent Gateway and Agent Registry. The focus is managing identity, access and observability as agents operate on enterprise infrastructure. This is the provider’s description: announced features do not amount to a security guarantee for every project. Businesses need to check configuration, permissions and accountability in their own environment.

Google Cloud · infrastructure and orchestration, 30 October 2026 ↗

NVIDIA · A route to testing models locally

NVIDIA publishes a guide to running Nemotron-3.5 Lightning on DGX Spark with TensorRT Edge-LLM and making the model available through a local application interface. The workflow connects prototyping with embedded-device deployment. It is a technical tutorial with specific hardware and software requirements, not a ready-made service for any computer. The reported speed results are NVIDIA measurements under the stated conditions, rather than automatic predictions for every real-world application.

NVIDIA · models for edge applications, 30 October 2026 ↗

The desk’s view · Autonomy needs verification

The common thread, in the desk’s view, is the shift from capabilities to controls: knowing what data goes in, which actions are allowed and how to intervene. Research, incidents and commercial announcements provide different kinds of evidence. Reading them separately helps avoid confusing an experiment with a product, or a promise with a result.

Sources and documents

Document date or reference period: Selection for 1 November 2026: three documents dated 30 October and an Anthropic study dated October 2026. These are not all announcements from the Italian overnight period..

Desk selection and summary with AI assistance, based on the four linked original sources. Laboratory experiments are distinguished from observed incidents and vendor announcements. The conclusion is the desk’s interpretation.

Report an update or correction →