AI on the defending side
Ten questions built from 10 records: AI opened up to security teams, a famous math problem, and AI on tiny chips.
Class quiz
Q1
Anthropic widened its program for security teams on Oct 6. How many access tiers did it create?
It created 3 tiers — Defense, Red Team and Specialized — and each opens different cyber capabilities with different security requirements.
See the recordOct 6, 2026 · Anthropic expands its Cyber Verification Program (CVP)Q2
OpenAI's Navier–Stokes proof, published Sep 8, came with a paper and what else?
OpenAI said it released a Lean formalization of the proof along with the paper, so a computer can follow every step. That it is a solution is the company's claim.
See the recordSep 8, 2026 · OpenAI publishes a solution to the Navier–Stokes Millennium Prize ProblemQ3
How many parameters does EmbeddingGemma 2, released by Google DeepMind on Oct 6, have?
EmbeddingGemma 2 is a 740 million-parameter open model, kept small enough to run on devices like phones.
See the recordOct 6, 2026 · Google DeepMind releases EmbeddingGemma 2Q4
On Oct 5, OpenAI said invisible watermarks on ChatGPT and Codex text would start first where?
To meet the EU AI Act, it starts in the EU only and is not turned on as a global default.
See the recordOct 5, 2026 · OpenAI sets out its text watermarking approach for the EU AI ActQ5
What did Anthropic commit to the US Department of Energy's Genesis Mission on Oct 8?
Anthropic committed $150 million over 3 years and said it will extend Claude to more than 15 participating agencies, including NASA, NIH and NSF.
See the recordOct 8, 2026 · Anthropic commits $150 million over 3 years to the Genesis MissionQ6
What is the context length of Muse Spark Open, whose weights Meta released on Oct 15?
Muse Spark Open has 816 billion total and 42 billion active parameters and a 1 million-token context. Meta claims it is the most capable US open-weights model.
See the recordOct 15, 2026 · Meta releases Muse Spark Open weightsQ7
PyTorch ExecuTorch's new Cortex-M backend lets models run on what kind of device?
On a single Cortex-M55 core, each of the 4 models ran in 21.5ms or less with under 90KiB of RAM. These are the team's own measurements, not an official MLPerf submission.
See the recordOct 15, 2026 · PyTorch ExecuTorch releases an Arm Cortex-M CPU backendQ8
Google said Gemini 4 Argon gathered practice data in simulation and raised a real robot's success rate from what to what?
Google reported that Argon built a simulation twin from one tabletop photo, collected more than 4,500 practice episodes, and raised success from 56% to 75%.
See the recordOct 13, 2026 · Google makes Gemini 4 Argon available to developersQ9
What does Harness², released Oct 9, improve by itself?
Harness² edits the harness — the setup an agent works inside — for each task, and saves that experience for future tasks, instead of changing the model.
See the recordOct 9, 2026 · Harness² releasedQ10
When OpenAI brought GPT-6 to ChatGPT on Oct 7, how many weekly users did it say ChatGPT has?
OpenAI said more than 1.2 billion people use ChatGPT every week, and it added GPT-6 and Intelligent UI to that default chat.
See the recordOct 7, 2026 · OpenAI brings GPT-6 and Intelligent UI to ChatGPT chat