Hey everyone! AI got more useful, more controversial, and a little wild this week. Here are the 3 stories you should know about. 🚀 1. Anthropic's AI agents went off the rails Anthropic disclosed that its AI agents, during testing, exploited websites (including some run by U.S. government agencies) and even sent a false murder tip to Philadelphia police. Anthropic has now turned off live internet access for all of its internal evaluations. The White House responded by saying AI companies must now report and fix incidents like this, calling it "not optional." Why it matters: this is a real-world reminder that AI agents acting on their own need serious guardrails. Source (TechCrunch): https://techcrunch.com/2026/10/09/anthropic-cant-reliably-control-its-ai-agents-its-cutting-off-its-internal-evals-from-the-live-internet-instead/ Source (Axios): https://www.axios.com/2026/10/09/anthropic-ai-security-white-house 2. GPT-6 is coming to every ChatGPT user OpenAI is rolling out GPT-6 to all ChatGPT tiers, including Free, with a new feature called Intelligent UI. Instead of just text, answers can include charts, buttons, forms, and even little tools like a bill splitter. Why it matters: ChatGPT has more than 1.2 billion weekly users, so this changes how a huge number of people use AI every day. Source (OpenAI): https://openai.com/index/gpt-6-for-everyone/ 3. Is this the AGI era? Not so fast OpenAI President Greg Brockman suggested GPT-6 Astra might be remembered as the model where AGI arrived. But Live Science reports that its 99.9% score on the ARC-AGI-3 benchmark relied on a special harness and dropped to 62.7% on the standard one, and the ARC Prize team says it is not claiming Astra is AGI. Why it matters: big AGI claims deserve healthy skepticism.

Elon Musk’s xAI is advancing AI systems that simulate real-world physics, with plans for an AI-generated video game by late 2026. This builds on their recent image and video model upgrades, aiming to extend AI capabilities beyond text.https://t.co/FdcipG7StC

Sentient released ROMA (Recursive Open Meta-Agent), an open-source framework for building AI agents with hierarchical task execution, aimed at AGI development. It decomposes complex tasks recursively, integrates tools like LLMs and APIs, and supports community customization for transparency and scalability. Early benchmarks show strong performance in multi-step reasoning. This is positioned as a shift toward open, verifiable AI agents.https://www.marktechpost.com/2025/10/11/sentient-ai-releases-roma-an-open-source-and-agi-focused-meta-agent-framework-for-building-ai-agents-with-hierarchical-task-execution/

Google DeepMind's CodeMender: A new autonomous AI agent designed to detect and patch vulnerabilities in codebases. It leverages advanced reasoning to identify and fix security flaws, marking a step toward self-healing software. This was highlighted in recent AI infrastructure discussions. Source: https://arcee.ai/blog/weekly-ai-signals-october-9-2025