Meta Publishes Six Math Papers Written With Muse Spark -- Using the Ordinary Chat App, Not a Research Harness
On October 2 Meta released six papers in which mathematicians worked with Muse Spark 1.1 and 1.2 in Thinking Mode through the standard meta.ai chat interface. Meta says five answer previously open questions, marks which passages the model drafted, and concedes that independent teams reached some of the same results at about the same time.
A Mythos-Found Bug Was Exploited the Day After Its Write-Up -- the Same Week Google Stopped Taking AI-Flooded Open-Source Bug Reports
Horizon3 published how Anthropic's Mythos found and weaponized CVE-2026-61500 in Rejetto HFS on September 30; VulnCheck saw exploitation from a China-based actor the next evening. On October 1, Google paused product-vulnerability submissions to its open-source bug bounty because most automated reports were invalid. AI bug-finding is getting better and worse at once.
WIRED: Meta's Muse Keeps 'a Page for Every Person in the User's Life' -- Its Own Instructions Show How
An October 3 WIRED report, based on internal files a researcher got Muse to share through ordinary chat, shows Meta's agent compiling hourly relationship pages on family, partners, friends, colleagues and people the user follows. Meta says the files were meant to be visible and that Muse uses only public information and what users share.
GitHub Trending: Impeccable Gives Coding Agents a Design Linter, and antirez's DwarfStar Runs DeepSeek V4 Locally
Two repositories on GitHub's October 5 daily list show what developers are building around agents: Impeccable, +1,171 stars, adds 61 deterministic checks for the visual 'tells' of AI-generated interfaces; antirez's DwarfStar (ds4) is a narrow, MIT-licensed inference engine for running DeepSeek V4 and a few other models on Macs, DGX Spark and Strix Halo machines.
Framework Watch: Google ADK 2.11 and OpenAI's Agents SDK 0.23 Both Ship on October 1 -- With Cancellation, Approvals and Redaction
ADK 2.11.0 adds graceful run cancellation, tool confirmation inside workflows, a budget-capped tool for consulting a second model, SQLite memory and an MCP SDK 2.x path. OpenAI Agents SDK 0.23.0 scopes tool approvals to the agent that owns the tool and redacts tool-failure details by default.
Trend Watch: Altman Says the World 'Should Accept Some Bad Things Happening' for AI's Benefits -- and Sees Daylight With Anthropic
In an October 4 Politico interview, Sam Altman said society should tolerate some harms in exchange for broad access to AI, called it 'a completely unacceptable trade-off' to let one lab decide nothing bad happens, and said OpenAI and Anthropic differ on risk -- weeks after agreeing with Dario Amodei's call to pace the frontier.