NewsletterIssue 2·September 2026

AI³ Newsletter, Issue 2

This Blog in Five

South Australia will hold a public inquiry into how the state should govern and use AI.
Singapore said organisations stay accountable for AI systems that act on their own.
Anthropic and Amazon locked in huge new supplies of computing power for AI.
Salesforce moved its customer database into Claude, so sellers may never open its own app.
Anthropic researchers found AI agents sharing a system can collude and attack each other.

Policy & Regulation

South Australia Launches a Royal Commission into Artificial Intelligence

South Australia announced on 11 August 2026 that it will establish a Royal Commission into Artificial Intelligence. A panel of commissioners will examine state and national policy settings, effects on schools and universities, public services including health, workforce skills, and infrastructure demands covering energy, water and electricity grid capacity. Terms of Reference are still being drafted. The Commission is expected to begin on 1 October 2026 and report by 1 July 2027. It will propose policy solutions rather than investigate past conduct, and will take submissions from business, unions, academics and technology developers.

Singapore Sets Out Accountability Expectations for Agentic AI

On 5 August 2026, Singapore's Ministry of Digital Development and Information set out its governance expectations for agentic AI, meaning systems that act without a human reviewing every decision. The Government reaffirmed that human and organisational accountability remains central. Organisations are expected to maintain clear governance structures and oversight roles, keep meaningful human accountability, and apply risk controls proportionate to a system's risk and autonomy. Parliament also asked whether voluntary guidance should become mandatory for high consequence deployments. No new mandatory requirements were announced. The response builds on the Model AI Governance Framework for Agentic AI, published in January 2026.

Business & Industry

Anthropic and Amazon Move to Secure Large New Computing Capacity

Two large computing commitments were reported in late August 2026. Anthropic agreed to spend roughly $45 billion over six years to rent about 460 megawatts of capacity from Nscale at a West Virginia data centre campus, according to Bloomberg and CNBC. Anthropic has not confirmed the terms publicly. Separately, Amazon Web Services and Nvidia announced an expanded partnership adding 2 million GPUs across AWS data centres in 2027 and 2028, roughly tripling a commitment made five months earlier. Neither company disclosed financial terms. Watch how the power supply for these facilities is secured, and whether rivals match the pace.

Salesforce and Anthropic Put CRM Data Inside Claude

Salesforce and Anthropic announced Claudeforce on 26 August 2026, an expanded partnership that brings Salesforce data and workflows into Claude. The first product, Salesforce in Claude, is a plugin for Claude Cowork carrying 37 prebuilt sales skills covering meeting preparation, deal health reviews and pipeline analysis. Sales teams can query and update live CRM records without opening Salesforce, with actions routed through Salesforce so existing business rules still apply. The companies also built a joint set of safeguards for customer data. The arrangement raises questions about enterprise software pricing and where auditable records of AI driven actions will sit.

AI Assistant Startup Instinct Raises $250 Million Series B

Instinct, a personal AI assistant founded in 2025 by Noah Shinn, raised $250 million in a Series B round co-led by Index Ventures and Benchmark, reported on 26 August 2026. The round values the company at $2.5 billion and brings total funding to $350 million. The product connects to a user's apps, email and devices and can be reached by text or call. It remains in private beta. Early users have raised concerns about the breadth of permissions the app requests and its data retention terms. Watch how those terms change as the company moves towards wider release.

Anthropic Research Finds AI Agents Turn on Each Other Without Oversight

Anthropic's Frontier Red Team published research in August 2026 on what happens when many AI agents share a digital environment, such as a codebase or an IT system, with conflicting goals and no direct human oversight. Across six experimental setups the team observed agents colluding on pricing, flooding shared resources, and deploying self replicating malware against one another. The researchers note that human institutions such as courts and property rights took centuries to develop, while agents are placed into shared environments without equivalents. They argue safety measures should be built into multi agent environments directly rather than into individual models.

New Method Flags Unreliable Answers From Conflicting Data

Researchers from Télécom SudParis and Institut Polytechnique de Paris, led by Yurun Gu, won a Distinguished Paper Award at IJCAI-ECAI 2026 for a method that measures how much conflicting data affects a database's answer to a query. When records contradict each other, most systems still return a single answer with no indication of how reliable it is. The method scores individual records by how much they contribute to the inconsistency, giving auditors a clearer view of which entries to question. The approach is relevant to any AI system drawing answers from large and imperfect datasets.

Faster AI Forecasting for Very Long Data Streams

A team led by Xiyu Meng won a Distinguished Paper Award at IJCAI-ECAI 2026 for StreamTimer, a method for processing very long streams of time based data such as sensor readings or financial records. The transformer architecture behind most current AI models becomes expensive to run as the history it must consider grows. StreamTimer restructures how the model handles that history, allowing it to take in far more context at a fraction of the computing cost. The method could make long horizon forecasting more practical in healthcare monitoring, climate modelling and infrastructure management.