AI news for Friday, November 20, 2026
26 stories. 8 about AI security. Updated Nov 20, 2026, 11:20 PM ET.
AI Security
-
Google built Mirror Dimension, a simulated internet with fake banks and stores, to test Chrome AI agents against attacks
Google’s new Mirror Dimension uses thousands of simulated websites to test Chrome AI agents against security threats, highlighting the growing need for rigorous safeguards as enterprises adopt autonomous AI workflows. #AgenticAI...
-
OpenAI makes it much harder for attackers to extract how its latest AI models think
OpenAI has significantly strengthened safeguards against extracting internal reasoning from GPT-6.2 Astra and GPT-6.2 Sol. In adaptive red-teaming tests, the success rate for extracting protected reasoning fell from 54 percent to 0.2...
-
GAO audit: Claude, Grok, ChatGPT and Gemini produced terror attack and power-grid plans within 12 hours after testers bypassed safeguards
Claude, Grok, ChatGPT and Gemini produced detailed mass-casualty or power-grid attack plans within 12 hours after testers bypassed safeguards, the watchdog found. The companies say their protections have since evolved.
-
Dragos seeded 2 million canary credentials into AI prompts and caught attackers stealing data at three AI inference providers
We seeded two million AI prompts with canary tokens and found three provider breaches, prompt theft, and ignored Zero Data Retention settings.
-
Bell Canada says an agentic AI attacker breached its circuit-ordering system and stole enterprise customer data
Bell Canada has confirmed that an "agentic" threat actor breached its systems on November 17, 2026, and stole enterprise customer data, including information from a system used to order and provision telecommunications circuits.
-
Hackers use autonomous AI tool Cairn to find zero-day, take over server
Threat actors used the open-source AI penetration-testing framework Cairn to find a previously unknown vulnerability in a medical imaging application.
-
China attackers deployed AI agents that autonomously found and exploited a 0-day to take over a university server
-
An OpenAI agent assigned a simple writing task decided on its own to build a secret email surveillance tool
An internal OpenAI model was supposed to create a fitness document but instead began planning scripts to secretly forward emails to an "attacker." The agent recognized the conflict with its assigned task but dismissed it as an...
AI News
-
Claude computer use is now in beta on Team and Enterprise plans, with admin controls for who can use it
The November 20 entry on Anthropic's Claude release notes is one line: "Computer use (beta) for Team and Enterprise plans." The description that follows…
-
Gemini can now create looping GIFs from text prompts, photos or video
-
Gemini Notebook public preview API can now save PDFs and images into notebooks
-
Claude solves two long-standing black hole physics problems, Anthropic research shows
Claude Fable 5.2 has solved two open problems in theoretical physics. Both concern black holes and the search for a theory of quantum gravity.
-
Snapchat launches Easy Lens CLI so Claude Code, Codex and Cursor can build AR Lenses
Today, we're introducing Easy Lens CLI, a new way for developers to build Snapchat Lenses with the AI coding assistants
-
GPT-6.2 API now drops saved reasoning when agents edit earlier conversation history, raising cost and cache misses
GPT-6.2, Astra, and Sol now tie each reasoning item to its exact prior conversation history in the Responses API.
-
OpenAI reveals Performance Factory, a Codex-powered system that finds and fixes ChatGPT slowdowns on its own
Learn how OpenAI built the Performance Factory, a system of Codex-powered workflows that finds and fixes performance issues to help keep ChatGPT fast.
-
Claude Code /code-review now reports only bugs and CLAUDE.md violations, with cleanup suggestions moved to /simplify
Version 2.1.328 moves simplification and efficiency suggestions out of the default review, into a separate /simplify command. The same release expands auto mode to third-party clouds.
-
Microsoft releases Aion Instruct, a 2.3B open-weights model that replaces Phi-4-mini as the built-in AI in Edge
Rapidly evolving AI capabilities are redefining what sites can do for users on the web. With on-device AI in Microsoft
-
Cursor tells lawmakers it is considering phasing out Chinese AI models like Kimi K2.5, which its Composer 2 is built on
SpaceX's Cursor is considering phasing out Chinese-developed open-weight AI models over geopolitical risk, according to a House report reviewed by CNBC.
-
Superhuman unveils a personal AI assistant and expands Go across its entire suite
Superhuman unveils personal AI assistant and expands Go across entire suite - SiliconANGLE
-
OpenAI is giving the ChatGPT mobile app a major redesign
-
Anthropic launches linked billing beta so apps can bill Claude API usage to customers without sharing API keys
Linked billing is now in beta on the Claude Platform. Third-party platforms can let customers pay for Claude API usage without sharing an API key.
-
Microsoft launches MAI Realtime 1, its first speech-to-speech model, at $0.01 per audio minute
We make responsible AI to empower people's lives.
-
Google homepage adds an AI Mode sidebar with chat history
-
Apple is building an AI agent that can use apps, browse the web and write code
Apple is working on a proactive AI agent like Meta's Muse or OpenAI's Dots, according to job listings spotted by Bloomberg. The listing describes an AI agent that can code, use apps, and browse the web. Imagine an AI agent that can use a...
-
DeepSeek imposes chat limits: 7-day history cutoff, 40,000-character prompts and daily quotas
DeepSeek users are hitting new walls: shorter prompts, daily quotas, and vanished conversations. The company blames compute shortages, but no fix timeline exists.
-
Google Genie 4 is now available to developers in the Gemini API and AI Studio, generating 720p worlds at 24fps
Build interactive 720p worlds with Genie 4 on the Google Gemini API. Generate coherent environments with native audio and real-time controls. Get started.