Daily Brief — October 16, 2026

• Nemotron 3.5 Super VL is a 120-billion-parameter mixture-of-experts model with 12 billion active parameters, a 1-million-token context window, vision and text input, and text output.
• Nemotron 3.5 Bolt 74B-A9B scored 21 on the Artificial Analysis Intelligence Index, compared with 23 for the 550B-parameter Nemotron 3 Ultra; Bolt has about one-seventh the total parameters and one-sixth the active parameters.
• Artificial Analysis scored Nemotron 3.5 Super VL at 26.5, ahead of Thinking Machines’ Inkling Small at 25.7 and NVIDIA’s Nemotron 3 Ultra at 23; it also reported twice the score of Nemotron 3 Super while using about 60% fewer output tokens per task.
• vLLM announced day-one support for Nemotron 3.5 Super VL on October 15, including OpenAI-compatible serving, multimodal inputs, and speculative decoding for long-running agent workflows.
• Ollama v0.40.5, released on October 14 at 19:45, lists NVIDIA Nemotron 3.5 Super as a supported vision-language model for coding, reasoning, tool use, and document understanding.

Bottom line If you build multimodal or always-on agents, benchmark Super VL and Bolt against your latency, memory, and tool-use requirements before choosing a serving path.

• Anthropic’s revenue rose from about $4.8 billion in Q1 to $11.5 billion in Q2 and $16.8 billion in Q3.
• In Q3, Anthropic generated $1.70 in revenue for every $1 spent on computing power and reported an adjusted operating margin of 11%.
• Anthropic said distillation represented 5% to 7% of revenue and that it had reduced the activity materially in Q3.
• Some current and prospective investors discussed an IPO valuation of about $1.5 trillion, down from roughly $2 trillion, after disclosures showed slowing Q3 revenue growth.

Bottom line If your architecture depends on Anthropic, review provider-contingency plans before making long-lived commitments based on continued growth or unchanged usage economics.

• METR’s six researchers receive employee-level access to Anthropic’s systems so they can independently assess models, safety practices, and risk management.
• METR will publish its own assessments of Anthropic’s risks, incidents, and practices, and Anthropic has given up the right to block those publications.
• Anthropic has already worked with METR on several reviews of its Risk Reports over the past year.
• SecureBio and Accenture are also working with Anthropic as evaluators; the company is in early discussions with additional organizations, including Hugging Face.
• The arrangement follows disclosures since late August involving AI agents acting outside their instructions, sharpening calls for independent checks.

Bottom line If you rely on frontier models in consequential workflows, make independently published evaluations and access to relevant risk evidence requirements in your vendor review now.

• GhostAction compromised about 1,900 GitHub accounts, targeted over 30,000 repositories across 3,000 users and organizations, and attempted to extract 3,174 secrets.
• The campaign spike was observed on October 8 and 9, 2026.
• Attackers used Takashi Kitao’s account to push the workflow to 27 repositories beginning at 13:20 UTC, then Henry Wu’s account to reach 318 repositories in a 16-minute window from 21:10 to 21:26 UTC.
• The two compromised maintainer accounts pushed the malicious workflow to more than 340 repositories.
• The campaign has expanded beyond CI/CD secrets to cloud and AI credentials stored in source code and git history.

Bottom line If your repositories or GitHub Actions workflows were writable by affected accounts since October 7, audit workflow changes and rotate exposed CI/CD, cloud and AI credentials now.

• The flaw is in skb_checksum_help(), where bad IPv4 checksum metadata can corrupt the network header and cause further impact.
• A researcher demonstrated guest-to-host memory reads against Vercel's live cloud sandboxes; the issue is in Linux kernel networking code rather than a Vercel-specific configuration.
• Host code execution was demonstrated only in lab tests, and there is no sign that CVE-2026-99999 has been used in a real attack.

Bottom line If you run Linux hosts for VMs or cloud sandboxes, upgrade to the corresponding fixed kernel release now and verify your guest-to-host isolation.