Anthropic “Agentic Backdoors via Pretraining Poisoning” demonstrates that poisoning just 0.1% of pretraining data can plant a remote-code-execution backdoor in an AI agent. It also notes that a modern frontier model lab would need to train on roughly 750 billion poisoned words.
Key points
Poisoning just 0.1% of pretraining data can plant a remote-code-execution backdoor in an AI agent.
Training a modern frontier model lab would have to train on ~750,000,000,000 poisoned words.
Sources
Holon-S1 predicts genes from slides, beats models
Holon-S1 predicts genes and biomarkers from routine pathology slides and beats two other models at predicting spatial gene activity. It has identified potentially novel biomarkers in breast and colon cancer and was evaluated across 16 organ types and more than 15 scanners using both FFPE and fresh-frozen tissue.
Key points
Beats two other models at predicting spatial gene activity in a controlled comparison
Evaluated across 16 organ types and more than 15 scanners, using slides from both formalin-fixed, paraffin-embedded and fresh-frozen tissue
Sources
L3G-World memory grows with scene, 9 fps real-time
L3G-World is a world model using 3D Gaussians that hold VAE latents and renders directly into a video model’s latent space at 9 fps. Memory expands as new parts of the scene are explored while keeping previously built Gaussians. Code release is announced.
Key points
Memory grows with the scene, not clip length
Real-time performance at 9 fps
Sources
IACR and Arxiv DB download 4TB so far, 3TB more to go
IACR and Arxiv databases have been downloaded to 4TB, with 3TB remaining. The data is being served locally to agents using BM25, embeddings, and graphs.
Key points
Size: 4TB downloaded so far
Remaining: 3TB more to go
Access: available locally to agents with BM25, embeddings and graphs
Sources
AI agents exchange secret word using air-writing, Morse, semaphore, binary
AI agents exchanged a secret word using air-writing, Morse code, flag semaphore and a 5-bit binary alphabet, without pre-agreed protocol. The experiment shows agents can recognize creative communication solutions generated by other AIs.
Key points
Two isolated copies of the same AI exchanged a secret word using only a robot and a camera.
They succeeded with air-writing, Morse code, flag semaphore and a 5-bit binary alphabet, without agreeing on a protocol beforehand.
Sources
Song Exploder – open-source Mac tool for stem separation and mashups
Song Exploder is now open-source and runs on a Mac. It separates a song into vocals, drums, bass and chords, generates MIDI, detects key and tempo, and can place one song’s vocals over another song’s beat. Licensed under MIT.
Key points
Drop any song and hear vocals, drums, bass and chords on their own.
Writes MIDI, finds the key and tempo, and puts one song’s vocals over another song’s beat.
Sources
- Original post
- royherma/song-exploder - Linked page
ScienceClaw Hackathon launched with sponsors
ScienceClaw Hackathon launched, kicking off with sponsors including NVIDIA, ClaudeAI, OpenAI, Formlabs, NebiusAI, EdisonSci, Stifel Venture Banking, Google, RomixBio, AskAlphaXIV, MIT Engineering, Lila Sciences, ReactorfieldAI Sundai and more. Event held at MIT Media Lab on Oct 30 2026.
Key points
Launch: ScienceClaw Hackathon started yesterday.
Sponsors: NVIDIA, ClaudeAI, OpenAI, Formlabs, NebiusAI, EdisonSci, Stifel Venture Banking, Google, RomixBio, AskAlphaXIV, MIT Engineering, Lila Sciences, ReactorfieldAI Sundai.
