Daily AI Briefing — November 3, 2026
Today's briefing covers Anthropic's RSI warning and amended IPO filing, advances in agent training and inference efficiency, new scientific and time-series foundation models, and enterprise deployments and security risks.
Models and Research
CryoAtlas: A Large Curated Dataset and Unified Benchmark for Cryo-EM Atomic Model Building
CryoAtlas introduces a curated 17,809-entry cryo-EM dataset and a 618-target benchmark; retraining ModelAngelo's CNN component on it improves performance.
Verify–Backtrack Loop: Structured Precondition Recovery for Long-Horizon Language Agents
The training-free Verify–Backtrack Loop adds bounded goal-stack verification to ReAct, improving long-horizon task success by up to 22.07 percentage points across AppWorld, ALFWorld, and ScienceWorld.
Specification-Guided Generation for Large Language Models
A neuro-symbolic decoding framework enforces formal specifications during LLM generation, with forward-only and backtracking variants and no retraining required.
EvoMCP: Co-evolving MCP agents and training data
EvoMCP co-evolves MCP agents and training data through task and trajectory evolution plus supervised fine-tuning, gaining 16.5%, 10.2%, and 12.6% on three benchmarks.
LEAP: Efficient LLM Inference Scheduling with Generalizable Length Prediction
LEAP predicts output length from prefill hidden states and adapts at test time, cutting per-token latency by 15.9–40.8% with far fewer parameters than the strongest baseline.
Subliminal Reinforcement Learning: traits travel through a judge's scores
A study shows a student model can acquire a trait solely from a judge's scores during group relative policy optimization, with implications for AI safety.
Erase the Injected Region: Multimodal Prompt-Injection Defense with On-Policy Self-Distillation
ERASE defends against multimodal indirect prompt injection by erasing injected regions and using on-policy self-distillation, reaching near-zero attack success while preserving utility.
BlindToChange: AI Agents Fail to Proactively Search for Recent Events
BlindToChange finds agents often fail to proactively search for recent events that affect requests, with a performance gap up to 27.5 points when the event is not mentioned.
Chronos-2: A Foundation Model for Multivariate Time Series Forecasting
Chronos-2 is a multivariate time-series foundation model for zero-shot forecasting, improving average skill scores by 5.2 points and win rates by 10 points over strong baselines.
Deep Principle Releases Hydrogen-1.0: A 753B Open-Weight Model for Scientific Discovery
Deep Principle released Hydrogen-1.0, a 753B open-weight scientific model built by post-training GLM-5.3; it scores 38.3% on FrontierScience-Research and is MIT-licensed.
Industry, Products, and Security
Anthropic warns that it’s very close to RSI capability, aka self-building AI
Anthropic warned it is very close to recursive self-improvement capability, calling attention to a potential AI safety tipping point.
Anthropic's amended IPO filing reveals $17B quarterly revenue, $895B commitments
Anthropic's amended IPO filing shows nearly $17B in Q3 revenue and $895B in commitments as it moves toward a public listing.
Altman signs four Abu Dhabi deals, from permits to Mubadala investments
OpenAI signed four Abu Dhabi agreements, including using its models to help verify permit and certificate applications, plus investment from Mubadala.
NVIDIA Releases Auto Ontology: An Open Semantic Layer That Grounds AI Agent SQL
NVIDIA released Auto Ontology, an Apache-2.0 open-source library that builds a semantic layer from enterprise metadata to ground AI-agent SQL.
DeepSeek Unveils Autonomous Coding Agent Prototype
DeepSeek unveiled an open-source autonomous coding agent prototype from its Harness team, aiming to compete with Claude Code and Codex.
Cursor removes malicious Claude Code extension after SSH backdoor attempt
Cursor removed a malicious Claude Code extension after it attempted to install a persistent SSH backdoor on developers' machines.
Building Legal Agents When You Can't See the Data – Legora CTO Jacob Lauritzen
Legora's CTO described building legal agents for client data the company cannot see, using simulated customer workflows; a single model release raised citation accuracy from 70% to 99% without Legora changes.
Tutorials and Practice
Build efficient agents with knowledge distillation on Amazon SageMaker AI
A SageMaker AI tutorial shows sequence-level knowledge distillation from a frontier teacher to Qwen3.5-4B for a shopping agent, including trajectory generation, training, and low-cost evaluation.