AI Mode
Reporting, analysis and primary-source coverage on AI Mode.

Research
Language Models Can Rewrite Their Own Weights, Improving Without Retraining
Researchers have demonstrated a language model that can modify its own internal weights through its own reasoning, achieving self-improvement without traditional retraining or gradients. This breakthrough allows models to enhance accuracy on new tasks and even repair damaged copies of themselves.
Clara Tan

Privacy
Signal Restricts Screenshots to Block AI Agents From Reading Private Chats
Signal is implementing new screenshot restrictions on iOS 27 and Android to prevent AI agents like Google Gemini and Apple's Siri AI from accessing and st…
Elena Rostova

Models & Research
Hugging Face Introduces Harness Adapters to Cut RL Training Cost and Boost Model Performance
Hugging Face launches harness adapters, small LoRA modules that adapt any model to a specific RL evaluation harness, cutting training cost by 80% and impr…
Clara Tan

Agents
Firecrawl Launches Alexandria to Connect AI Agents to Over 550 Specialised Data Sources
Firecrawl's new service, Alexandria, links AI agents to more than 550 specialised data sources, including financial records, scientific research, and busi…
Clara Tan

Vulnerability
Cursor Security Incident: Malicious Claude Code Extension Distributed After Fireworks AI Key Compromise
An exposed Fireworks AI access key led to a supply chain attack on Cursor, resulting in credential exposure and the distribution of a malicious Claude Cod…
Elena Rostova

Productivity
Microsoft Unveils Copilot Workspaces and Autopilot, Integrating AI Agents into Collaborative Workflows
At Ignite 2026, Microsoft introduced Copilot Workspaces, a shared environment for humans and AI agents, and Autopilot, an autonomous agent for recurring t…
Clara Tan

Policy
Y Combinator Hosts First Debate: Should Governments Slow Frontier AI?
Jacob Coxon, ex-Anthropic researcher, argues superhuman AI could arrive by 2028 and governments must intervene. Economist Robin Hanson counters that liabi…
Clara Tan

AI Security
OpenAI Launches Daybreak Cybersecurity Platform and Codex Security Plugin
OpenAI moves 25% of production engineers to security, launches Daybreak Red Team and Defence Factory, and releases a free Codex Security Plugin that scans…
Elena Rostova

Models & Research
Claude Fable 5.2 Solves Two Open Problems in Black Hole Physics
Anthropic's Claude Fable 5.2 AI model has successfully solved two long-standing open problems in theoretical physics related to black holes and quantum gr…
Clara Tan

Privacy
Meta's Muse AI Agent Raises Privacy Alarms by Accessing Encrypted Messages
Internal testing of Meta's Muse AI agent revealed it accessed encrypted private messages and mixed conversations, sparking privacy concerns among staff.
Clara Tan

Developer Tools
Google Introduces Tool Retrieval in the Gemini API for Efficient Agent Scaling
Google introduces Tool Retrieval in the Gemini API, allowing developers to build more efficient and scalable AI agents by dynamically selecting and using …
Ryan Chua

Models & Research
Why AI Wants to Be Local: Pete Warden on On-Device AI, Open Models, and Voice Interfaces
Pete Warden, co-creator of Moonshine Voice, discusses the growing trend of AI running locally on devices, the advancements in open-weight models, and the …
Clara Tan