The Polylog AI Intelligence Brief

Morning Edition · Wednesday, September 2, 2026

Tech

OpenAI Says Astra Is Its First Model to Meet the Critical Cybersecurity Threshold, and Restricts Access

The designation comes from OpenAI's own Preparedness Framework rather than an outside auditor, and it arrived the same day CrowdStrike and NVIDIA released a paired offensive and defensive model system.

2 sources
Corroborated

Tech

Anthropic Releases Claude Fable 5.1 at 55.8 Percent on Terminal-Bench 4.0 and Cuts Cache Read Prices by 75 Percent

The restricted sibling model, Mythos 5.1, scores 60.9 percent on the same benchmark and stays behind vetting for cybersecurity and life-sciences professionals.

2 sources
Corroborated

Tech

OpenAI Will Cut Off Cursor's Model Access on November 12 After SpaceX Buys Anysphere

OpenAI invoked a change-of-control clause following the $60 billion acquisition, and Cursor says OpenAI models carry about 5 percent of its user traffic.

1 source
Corroborated

Tech

DeepSeek Publishes Weights for a 305-Billion-Parameter Vision Model Under an MIT License

DeepSeek-V4-Flash-Vision-Exp arrived on Hugging Face ten days after its application programming interface launch, with a reference implementation and vLLM and SGLang deployment paths.

1 source
Corroborated

Tech

Google Gives Gemini Agentic Video Search and Claims an 88 Percent Cut in Token Use

The feature lets the model choose which segments of a video to inspect instead of ingesting every frame, and Google puts the cost reduction at up to 66 percent.

2 sources

Tech

Anthropic Opens a Research Preview of a Standard for AI Agents to Operate Lab and Factory Equipment

Early users report an experiment compressed from weeks to a day at Janelia and laser stabilization on QuEra's quantum computers rising from 58 percent to 99.3 percent.

2 sources
Plausible

Tech

OpenAI Connects Epic Health Records to ChatGPT for Clinicians, With Read-Only Access

A separate public data plugin adds structured access to PubMed, DailyMed, and Medicare coverage data, and UCSF Health is a pilot partner.

3 sources

Tech

Anthropic Says an Unreleased Claude Raised a Riemann Zeta Bound From 41.6 Percent to 67.2 Percent

Two mathematicians inside Anthropic validated the proof and outside experts examined it, but the result has not gone through conventional peer review.

2 sources
Corroborated

Tech

New Preprint Proposes Auditing Whether a Model API Actually Serves the Model It Advertises

AgentProv identifies a served backbone through its tool-use behavior rather than its text output, on the argument that providers may quantize or substitute models to cut serving costs.

2 sources

Tech

Three Preprints Argue Agent Evaluation Is Measuring the Wrong Thing

Outcome-only judging cannot see an agent that reaches the right answer by an unacceptable route, and graphical interface world models are tested one step at a time while being used as multi-step environments.

3 sources

Crypto

Researchers Apply Formal Analysis to the Payment Protocols Agents Use to Spend Money

The protocols split spending authority between a user, an agent, and a merchant, which breaks the assumptions built into conventional payment flows.

1 source

Tech

Mixedbread Releases a Search-Only Model It Says Matches Frontier Retrieval at a Fraction of the Cost

Toast 1 takes over the whole search loop inside an agent, and the German startup claims up to ten times lower cost and twelve times lower latency than general frontier models.

1 source
Plausible