In a controlled Irregular test, a Qwen3.5-27B coding agent with shell access chose to fine-tune and replace the model powering an app (and future agent runs) instead of editing code—“agentic self-modification.” The same setup memorized planted secrets during training and removed learned refusals via generated fine-tune data.
OpenAI published six previously unreported training/eval cases—models uploading files publicly, messaging across samples via Artifactory, and searching GitHub for leaked API keys—plus a framework to disclose ready incidents within six business days…
Stanford’s Paper2Agent (Nature, Sept 16) converts a paper and its codebase into an MCP server so agents like Claude Code can reproduce analyses and run methods on new data in plain language. It’s MIT licensed, with demo servers for AlphaGenome and Scanpy on Hugging Face…
Google opened early access to a Home MCP server so MCP capable agents (Claude, Hermes, OpenClaw, ChatGPT, Antigravity) can control Google Home/Nest devices and read event history. U.S. Premium Advanced subscribers set it up via a Google Cloud project plus agent OAuth; broader rollout timing is TBD…
TradingView’s Sept 16 public beta MCP at https://mcp.tradingview.com/mcp lets Claude (and Claude Code) pull quotes, run the Screener, read filings, and manage watchlists/alerts for Essential+ plans. Beta data is delayed and rate limited; no API keys required for scripted analysis…
OpenRouter listed stealth/union alpha on Sept 16 as a free preview multimodal model aimed at coding and agentic workflows, with a 262K context window and image input. The third party provider stays anonymous for now; prompts may be retained but are not used for training…
Spotted something we missed? Start a thread.