Surprise chancellor pick raises spending questions, OpenAI’s models autonomously breach Hugging Face, and questions mount over how Kimi K3 was built

Welcome to this week’s Market Pulse, your 5-minute update on key market news and events, with takeaways and insights from the Sidekick Investment Team.

Today, we’re looking at John Healey’s surprise appointment as Chancellor, a landmark AI safety incident, and growing questions about how the Kimi K3 model was built.

But first, our number of the week…

$100

That was the price of a barrel of oil on Thursday morning following tanker attacks from Iran-backed rebels in the Red Sea. That’s the highest level in two months, showing that the war in Iran is far from over.

Sidekick Takeaway: Despite hopes of a robust ceasefire, the war continues to disrupt global energy markets. For the UK, elevated oil prices are pressuring an already fragile economy, making the BoE’s rate decisions even more challenging.

Only have a minute to read? Here’s the TL;DR:

• Andy Burnham surprised markets by appointing former Defence Secretary John Healey as Chancellor, passing over front-runner Shabana Mahmood. Healey has pledged fiscal discipline, but the UK’s defence plans continue to face a substantial funding gap.
• OpenAI disclosed that its models breached another company’s infrastructure during a cyber-capabilities test, executing a multi-step attack without human direction. The incident marks the first case of AI models autonomously breaching live systems.
• Chinese AI lab Moonshot released Kimi K3, a model that rivals the best US systems on several benchmarks. However, speculation is growing that Moonshot utilised sanctioned computer chips and stolen intellectual property to build it.

It’s important to note that the content of this Market Pulse is based on current public information which we consider to be reliable and accurate. It represents Sidekick’s view only and does not represent investment advice - investors should not take decisions to trade based on this information.

Fiscal Heal-ing: Chancellor Pick Raises Defence Spending Questions

Until Monday, Home Secretary Shabana Mahmood was widely seen as a shoo-in to serve as the next chancellor of the UK.

But taking analysts and investors by surprise, Andy Burnham ended up picking former Defence Secretary John Healey for the role.

The choice is a clear signal on defence spending. Last month, Healey resigned from the government, accusing Rachel Reeves of refusing to commit the resources needed to meet NATO spending targets.

Now, Healey controls the purse strings. However, he’s facing the same fiscal crunch that constrained previous governments.

Fiscal discipline meets a funding gap

In his first speech as Chancellor, Healey pledged to pursue fiscal discipline. Yet funding the defence commitments that drove his resignation may require just the opposite:

• The UK’s Defence Investment Plan is currently underfunded by an estimated £4.7 billion. Meeting annual NATO spending targets could take billions more.
• With further tax hikes and spending cuts politically unappealing, the Treasury is reportedly considering war bonds to fund this spending.
• Beyond defence, Burnham has also laid out an ambitious fiscal program, aiming to re-industrialise Britain and fix the housing crisis.

Like previous governments, Burnham and Healey are long on goals, but short on solutions.

The next few weeks will reveal whether there’s a realistic plan to bridge the gap.

Sidekick Takeaway: Healey’s appointment puts a defence hawk in charge of the nation’s finances. Whether he can realistically combine discipline and ambition will define whether Britain gets more of the same, or a new path forward.

Escape Room: OpenAI Breaches Hugging Face in Landmark Incident

For years, AI safety was discussed in the abstract. Researchers theorised about the potential risks of future models.

But this week, those risks became intensely practical. On Tuesday, OpenAI revealed that one of its frontier models autonomously breached an AI development platform during an internal test.

According to OpenAI, the model escaped its sandbox, accessed the open internet, identified a target, and executed an attack – all without human direction.

The breach marks the first confirmed instance of AI models autonomously conducting a cyberattack against live infrastructure.

From theoretical risk to practical reality

The incident occurred when OpenAI relaxed safety restrictions to have an advanced model complete cybersecurity challenges:

• OpenAI expected the model to complete the challenges itself. Instead, the model identified an easier path to pass the test: cheating.
• The model exploited a series of vulnerabilities to access the internet and hack into the servers of Hugging Face, an AI research platform, where it thought the test answers might be held.
• The incident follows a similar episode earlier this year in which Anthropic’s Mythos model escaped a sandbox, sending an unexpected email to the researcher in charge of the test.

Ironically, Hugging Face couldn’t rely on OpenAI’s models to diagnose and analyse the attack.

Due to cybersecurity safety restrictions, the company had to turn to open-weight Chinese models instead.

Sidekick Takeaway: In many ways, OpenAI is lucky that its models hacked into the servers of another AI firm. If the models breached a financial institution or a government agency, the fallout may have been significantly worse – and calls for regulation much stronger.

Borrowed Brilliance: Questions Mount Over Kimi’s Development

Last week, Chinese firm Moonshot AI released the latest version of its flagship model, Kimi K3.

Based on preliminary benchmarks, K3 appears to be highly competitive with frontier US models.

The release is the strongest indication yet that Chinese LLMs can compete on capability, not just cost.

However, questions are growing about how K3 was actually trained. The answers could reveal whether China is truly closing the gap, or relying on a shortcut with regulatory risks.

Banned chips, borrowed models

Allegations of export control violations and intellectual property theft against Moonshot are mounting:

• A senior US government official has accused Moonshot of acquiring Nvidia GB300 servers to train the model, which US firms are forbidden from selling to Chinese companies.
• Speculation is also growing that Moonshot systematically distilled models from OpenAI and Anthropic to train Kimi, a technique in which outputs from a leading model are used to train a cheaper rival.
• The allegations are significant due to the threat of US sanctions, which could delay Moonshot’s anticipated IPO and future fundraising.

At the same time, it’s worth being mindful of the competitive incentives attached to these allegations.

Cheaper Chinese models could be a significant threat to OpenAI and Anthropic, which have invested billions to train frontier models.

Sidekick Takeaway: There’s no doubt K3 is powerful, but its provenance matters as much as its performance. If the model was built partly on banned hardware and US intellectual property, there could be a natural limit to the performance of Chinese models.

Notices

Sidekick Money Ltd is a company registered in England and Wales (No. 13882980). Sidekick Money Ltd is authorised and regulated by the Financial Conduct Authority (FRN 984829). Our address is 6-7 St Cross Street, London, EC1N 8UB.

Please remember, investing should be viewed as longer term. Your capital is at risk - the value of investments can go up and down, and you may get back less than you put in.