Attacks on autopilot — how do defenders catch up?
Ten questions built from 33 records: from automated attacks to agents porting old code and an olympiad math result.
Class quiz
Q1
Which statement best matches the conclusion of Anthropic's report 'Detecting and countering misuse of AI: September 2026'?
Sorting the operations it disrupted from December 2025 to August 2026 into 7 types, Anthropic said AI has collapsed the gaps in labor and tools. One group even went after production API keys to reach unreleased Claude models, and failed.
See the recordSep 10, 2026 · Detecting and countering misuse of AI: September 2026Q2
NVIDIA says its system, which writes proofs purely in natural language, scored how much at IMO 2026?
The authors claim it scored 30 out of 42 at IMO 2026 without formal provers, external tools or internet access, reaching the gold-medal threshold. They released the checkpoints, training data and code as well.
See the recordSep 9, 2026 · NVIDIA analyzes post-training and test-time design for olympiad proof generationQ3
How much did Mistral AI raise in the Series D it announced on Sep 8?
Mistral AI raised €3 billion at a post-money valuation above €21 billion. Samsung Electronics was among the investors leading the round.
See the recordSep 8, 2026 · Mistral AI raises €3 billion Series D at a valuation above €21 billionQ4
In Q2 2026, which programming language did OpenAI use to rewrite Habitat, ChatGPT's storage platform?
Two engineers rewrote the whole thing in Rust with Codex and GPT-5.5, gaining 6 times the CPU efficiency and 15 times the memory efficiency of Python. Habitat serves more than 1 billion weekly users.
See the recordSep 11, 2026 · Scaling Storage for 1 Billion ChatGPT Users (Part I)Q5
When Mistral AI used agents to port old Fortran 77 code, what role did people play?
Mistral says it was a structured workflow, with planner, coder, tester and code quality reviewer roles and human review gates. In the first sprint, 40,000 of the 300,000 lines were ported to C++.
See the recordSep 9, 2026 · Mistral AI shares agent-led modernization of a reservoir simulator's legacy codeQ6
How does Skild AI say its robot model S1 learns a new task?
Skild AI says S1 picks up new tasks in context from a single demonstration, without weight updates or retraining, and handles multi-step tasks up to 10 minutes long. These are the company's claims.
See the recordSep 10, 2026 · Skild AI releases robot foundation model S1Q7
What kind of voice model is GPT-Live-1, which OpenAI released in its API?
GPT-Live-1 is a full-duplex voice model that listens and speaks at once, doing away with the old STT→LLM→TTS chain. The voice layer costs $0.05 per minute.
See the recordSep 10, 2026 · OpenAI releases GPT-Live-1 in the APIQ8
In which country did NVIDIA say it would work with 8 data center operators to build up to 2 gigawatts of AI factory capacity by 2027?
NVIDIA announced it will work with 8 Australian data center operators to build up to 2 gigawatts of AI factory capacity by 2027. The total investment was not disclosed.
See the recordSep 9, 2026 · NVIDIA and 8 Australian data center firms plan up to 2GW of AI factory capacityQ9
What does Cohere's North Small Translate 1.0 do?
It's a model specialized for machine translation, covering more than 50 languages, with weights released under CC BY-NC 4.0 (non-commercial, research use). Scores such as its 83.6 average on WMT26 are company-reported.
See the recordSep 10, 2026 · Cohere releases North Small Translate 1.0 for machine translationQ10
Paul Christiano, who joined the OpenAI Foundation board, is known for contributing to which research?
Christiano is a former head of alignment research at OpenAI (2017–2021) who contributed to foundational RLHF research and founded the Alignment Research Center. He now evaluates frontier models as a senior technical advisor at NIST's Center for AI Standards and Innovation.
See the recordSep 9, 2026 · Paul Christiano joins the OpenAI Foundation board