Tech news, 12 November 2026: NEAR AI and Fireworks AI breached, a web host entered through its MCP server, Rackspace signs GB300 capacity and Anthropic sets its IPO range

1. NEAR AI breached through a flaw in the SGLang inference engine

NEAR AI disclosed that an attacker used a remote code execution flaw in SGLang, a widely used open-source inference engine, to run code on one of its inference servers². The Cloud Security Alliance notes that the attacker could observe in-flight requests on that server, including ones meant to be protected by confidential computing, because the enclave protects against the host and operator but not against flaws in software running inside it¹. NEAR AI said it revoked credentials and deployed a fix across its inference servers².

CTO take: confidential computing is a strong control, but it does not patch the software inside the enclave. If you self-host inference with SGLang, vLLM or similar engines, treat them like any internet-facing service: keep them patched, never expose their internal ports, and restrict what the serving process can reach. If you buy confidential inference, ask providers for their patch cadence and egress controls, not just their attestation. See AI model supply chain security.

2. Fireworks AI discloses a second breach

Fireworks AI told customers that an October intrusion through a third-party cloud provider reached its development machines, exposing account and billing information and, for a limited set of accounts, secrets, datasets and incomplete model weights³. It is the company's second disclosed incident in weeks; it has restricted some data upload and download functions while it investigates³.

CTO take: any provider you send fine-tuning data or keys to is holding your crown jewels. Rotate any secrets stored with affected providers, review what datasets you have uploaded and whether they are still needed, and keep the ability to move inference to another provider at short notice. See model routing and the third-party access case study.

3. A web host is breached through its MCP server

German hosting provider Mittwald said an attacker exploited a flaw in its MCP server, which connects customers' AI assistants to its mStudio platform, and accessed 56 customer projects before being detected; it shut down the server, revoked session tokens and asked users to change passwords and review API keys⁴. Whether the flaw was in Mittwald's own code or an underlying component has not been confirmed⁴.

CTO take: an MCP server that can act on customer resources is a privileged API and needs the same rigour: strong authentication, narrow scopes per user, rate limits, and logging that would spot a stranger moving between customers. If you publish an MCP server, review it this week. See the MCP server security checklist and the Europe page.

4. Rackspace signs a GB300 capacity deal with Together AI

Rackspace announced a multi-year agreement to provide Together AI with dedicated Nvidia GB300 NVL72 infrastructure, valued at about $380 million and expected in service by mid-2027; Converge Digest notes it remains conditional on financing and on equipment, power and facility availability⁵.

CTO take: new inference capacity is being contracted now for 2027, with financing still to be arranged. If you depend on a specialist inference provider, ask where its future capacity is coming from and how firm those plans are. See AI capacity contracts.

5. Anthropic sets its IPO price range

Anthropic set a price range of $70 to $78 per share for its IPO, and people familiar with the plans told CNBC it aims to start trading on 19 November⁶.

CTO take: for customers, little changes on listing day. Over time, public-market pressure can shape pricing, product priorities and disclosure. Keep provider choice open and contracts flexible, as with any strategic supplier.

What I'd do this week

• Patch self-hosted inference engines and check their ports are not exposed.
• Rotate secrets held by AI providers that have disclosed incidents.
• Review any MCP servers you publish for authentication, scoping and logging.
• Ask inference providers about their capacity plans for 2027.

I am based in India, the United States and the UAE, working globally. See where I work, or analyse your challenge and bring the result to a call.