# AI news, Wednesday, 2 September 2026

Compiled by Richard Brooks's AI news desk. Canonical: https://richard-brooks.com/ai-news/2026-09-02/

## Anthropic improves alignment and security efforts

Security. Resumed external cyber evals, a real-time escape and probe classifier, and partner sandbox guidance. Affects how Claude agents are scoped and monitored.

Source: Anthropic, https://www.anthropic.com/news/improving-alignment-security-efforts

## DeepSeek publishes V4-Flash-Vision-Exp weights

Models. MIT-licensed ~305B multimodal MoE weights with vLLM and SGLang serve notes for screenshot, chart and document workflows.

Source: DeepSeek, https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-Vision-Exp

## Google Antigravity Teamwork posts multi-agent wins

Agents. Multi-agent Teamwork updates claimed open theory results verified in Lean, a RISC-V simulator that boots xv6, and upstream Eigen and ParlayHash optimisations.

Source: Google, https://blog.google/innovation-and-ai/technology/developers-tools/antigravity-teamwork-multi-agent/

## Salesforce open-sources ClaimProbe and ClaimWriter

Language. Claim-level audit cut hallucination about 2.6 to 4.5 times in deep-research hosts where rubric scores can hide claim failures.

Source: arXiv, https://arxiv.org/abs/2608.28643

## Sony and UCLA post AToM CoWriter writing support

Language. Infers writing processes from keystrokes and fires support agents so writers need not craft prompts.

Source: arXiv, https://arxiv.org/abs/2608.30424