Anthropic says it may be near recursive self-improvement threshold

Anthropic filing warns it may soon hit recursive self-improvement threshold

Anthropic’s revised S-1A filing, submitted Monday, says the company may imminently meet its Responsible Scaling Policy definition of dramatically accelerated progress attributable to AI, the threshold it uses for recursive self-improvement. The filing says Anthropic has already used AI to develop models whose capability leaps it did not recognize at the time, and that internal estimates about future models, not current ones, suggest the threshold may soon be met. No current model has been designated as exceeding it. Anthropic warns that using AI to accelerate its own research may outpace its ability to develop and validate safety, alignment and security measures, and that it has previously failed to identify significant acceleration until after it occurred. It says it is considering policy updates to keep humans in control of frontier model development.

Anthropic warns that it’s very close to RSI capability, aka self-building AI →

Topics: Anthropic

Vapor Panda used ARTEX AI agents to attack over 200 organizations

Chinese hackers used ARTEX, an open-source multi-agent AI penetration-testing framework, against more than 200 organizations in 14 countries, CrowdStrike reported. The campaign began October 17 and targeted government ministries, telecom providers, and pharmaceutical, semiconductor, defense, nuclear and financial sectors. The actor, tracked as Vapor Panda, compromised at least nine organizations and ran password spraying and credential scanning against 12 other high-value targets. Confirmed victims include telecom providers in the U.S. and Belgium, government entities in Greece, the Middle East and Latin America, and U.S. pharmaceutical firms. CrowdStrike said the campaign marks an escalation from single intrusions to multiple simultaneous AI-directed engagements, with operators instructing agents in Chinese and intervening at critical decision points. In one intrusion agents consumed over 1.4 billion input tokens.

Chinese hackers use ARTEX AI in automated attacks against 200 orgs →

Topics: Autonomous AI agent cyberattacks in the wildState-sponsored hacking

Goldman Sachs bars most traders and bankers from AI coding tools

Goldman Sachs has barred most traders and bankers from using AI coding tools, including Anthropic’s Claude Code and Microsoft’s GitHub, under internal guidance updated in October that limits them to engineers. The bank had allowed bankers to use the tools for months before a review of productivity and risk concerns prompted the change. The restriction frustrated some traders who had used them to follow client trading, simulate market shifts and draft text and visuals. It does not cover strats, employees combining software engineering, quantitative research and trading. A Goldman spokesperson said the bank makes certain AI tools available, then after experimentation puts them in the hands of those best suited to use them long term. Goldman has also built its own AI assistant using models from OpenAI, Anthropic and Alphabet.

Goldman Sachs is restricting traders and bankers from using AI coding tools →

Topics: Claude Code

Tsinghua study finds coding agents can cut costs 26% by rewriting their own harness

A Tsinghua University-led study, HarnessPro, found AI coding agents can rewrite their own harness, the layer defining tools, context assembly, verification and reasoning budget, to improve accuracy while cutting cost. The framework runs a four-agent evolution loop using ten training tasks per benchmark. On SWE-Bench Pro, its Heavy configuration reached 80.7% average pass rate versus 80.0% for RHO, at $1.97 per task versus $2.68, a 26.5% cost reduction, with execution time falling from 520 to 412 seconds. Across three benchmarks Heavy averaged a 2.0-point pass-rate gain and 26.1% cost reduction. The paper is titled HarnessPro: Recursive Harness Self-Improvement with Pareto Optimization for Performance-Cost Trade-offs, with a v1 arXiv submission dated October 19, 2026.

Tsinghua Study: AI Coding Agents Cut Costs 26% by Rewriting Their Own Scaffolding →

ChatGPT adds PIN-locked chats for private conversations

OpenAI introduced Locked Chats, a feature protecting individual ChatGPT conversations with a four-digit PIN, with Face ID, Touch ID or device unlock usable instead on supported devices. Locking a chat moves it to Projects and Locked chats, removing it from regular history, pinned chats and standard search results; its title and contents stay behind the PIN until opened. Locking removes a conversation from its original project, so it loses that project’s instructions, files and memory. The feature rolls out gradually to Free, Go, Plus, Pro, Business, Enterprise and Edu users on web, iPhone, Android and the new Mac and Windows desktop apps; ChatGPT for Healthcare does not support it. Locked Chats provide no end-to-end encryption, and resetting a forgotten PIN permanently deletes all locked chats.

ChatGPT now lets you lock private chats with a PIN →

Topics: ChatGPT