OpenCode key breach and Replit's switch to OpenAI top the week

Attackers reach provider keys stored in OpenCode Console

Anomaly Innovations, the company behind the open-source OpenCode coding agent (it was called SST until recently), told affected customers by email on 17 October that someone had gained unauthorised access to a legacy system, which it's now shutting down. The company says roughly 1.4% of OpenCode Console users were hit. Only accounts that had linked outside providers were involved, and the notice warned that OpenAI and Anthropic keys saved in those workspaces may have leaked. Google keys were mentioned too.

The firm linked the incident to "a large scale distillation attack" and says it's working through it with Anthropic, OpenAI and Google. It hasn't said whether anyone misused the keys, and it says it's found no sign that other data leaked. Some users, though, say they ran up bills in the hundreds of dollars for queries they never sent, in some cases before the notice reached them.

What it means for NZ organisations: If your developers use OpenCode Console, cancel every linked key and create new ones in each provider's own dashboard rather than through email links, then go through usage logs from 17 October onwards. Put hard spending caps on every API account so one stolen key can't run up an unlimited bill.

Replit puts OpenAI first for its coding agent

Speaking at TechCrunch Disrupt on Thursday 16 October, Replit's chief executive Amjad Masad said the company had stopped working with Anthropic, giving "cost, partnership, and performance" as the reasons. Replit had been a top-five Anthropic customer until recently. Later he clarified that customers can still pick Claude models, but OpenAI's models are now the default and OpenAI is the "preferred partner" for Replit Agent. Since most users don't change the default, that could cut what Replit spends with Anthropic by a lot. A person familiar with the arrangement told TechCrunch the Anthropic partnership hasn't ended.

Masad also put enterprise at about 15% of revenue this year, compared with roughly 2% last year, and said B2B makes up more than half the total.

What it means for NZ organisations: The model behind a coding platform can change with no notice to you. Teams using Replit should re-test the quality of what it produces, and procurement teams should ask vendors how they'll announce model changes.

METR staff get inside access at Anthropic

In a post dated 15 October, Anthropic said it has brought six METR employees on board with roughly the same hardware, desk space and internal system access as its own risk assessment teams. METR, a nonprofit evaluator, will publish what it finds about Anthropic's risk levels, incidents and practices, and Anthropic can't block publication. Anthropic can require redactions in categories agreed in advance, and it decides what falls into them. Customer data, legally privileged material and model architecture details are kept out. Separately, SecureBio is reviewing the chemical and biological weapons chapter of Anthropic's August Risk Report.

What it means for NZ organisations: When you assess an AI supplier, ask which outside reviewers can examine its systems and whether they're free to publish what they find.

Meta's Muse Spark Open model fits on two Mac Studios

Meta published the weights of Muse Spark Open on 15 October. It's a natively multimodal model with 816 billion parameters in total, 42 billion of which are active, and it can take up to 1 million tokens of context. It's built for coding and agent tasks. Meta says that, on its own tests, it's the strongest open-weight model from a US company so far. Two Apple M5 Ultra Mac Studios, giving about 1 TB of memory between them, are enough to run it. You can download the weights from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models, and hosted access is offered by Groq, Fireworks AI, Together AI, OpenRouter and others. Use is governed by the Muse Model License, which has limited commercial terms.

What it means for NZ organisations: Firms that must keep data on their own premises now have a large model they can host themselves. Get the licence reviewed before you build anything commercial on it.

DeepSeek updates its Flash model and opens web search

On 17 October DeepSeek moved its deepseek-flash endpoint to V4.1-Flash-1017. It reports 77.0 on DeepSWE v1.1 and 42.3 on Terminal-Bench 4.0, both up on the previous V4.1, and says tasks now run faster with shorter outputs. Prices and the way you call the API haven't changed. On the same day, developers got access to DeepSeek's own web search engine through a Search and Fetch API.

What it means for NZ organisations: Low-cost coding models keep putting pressure on prices. Before adopting one, check that data handled outside New Zealand fits your privacy and procurement rules.

US brokers shut out unauthorised AI agents

Charles Schwab CEO Rick Wurster said on Bloomberg TV on 16 October that the firm stops unapproved AI agents from transacting in client accounts. His weekly report shows "tens and twenties" of blocked attempts by major agents. Outside AI tools can connect through Model Context Protocol servers instead, with read-only access to balances and research and no ability to move money. JPMorgan's Jamie Dimon has said his bank also keeps outside agents away from customer account data.

What it means for NZ organisations: Expect local banks to set similar limits. Build agent projects on approved read-only connections, and never give an agent a customer's login.

Character.AI enters Chapter 11

Character.AI filed for bankruptcy on 13 October, with assets and liabilities each put at between $50 million and $100 million. Disney will pay $20 million for some assets and get a perpetual, non-exclusive licence to the technology. The deal should close in late November or early December. Chief executive Colin Adams said the service will shut down if no workable bid comes in for the rest. The company had faced lawsuits over teen safety and an investigation in Texas.

What it means for NZ organisations: Well-known AI services can still collapse. Keep regular exports of your data and an exit plan for every AI vendor you depend on.