LooksmaxxingOct 22, 20264 min read
Face Analysis AI Gains Edge Friendly 3D Gaussian Head Avatar Research
A 7 October 2026 preprint explores efficient 3D Gaussian head avatars for face analysis AI on phones and edge devices.
Cyber SecurityOct 22, 20264 min read
Long Term Agent Memory Can Extend Prompt Injection Harm Study Warns
A 20 October 2026 preprint warns that long term agent memory can turn one prompt injection into lasting compromise.
Supply ChainOct 22, 20264 min read
Columbia Team Studies Blocking Model Theft From Planning APIs
Researchers study defenses against model extraction from contextual linear optimization APIs used in routing and sourcing math.
Agri TechOct 22, 20264 min read
GRAFT Model Answers Tomato Leaf Disease Questions From Photos
GRAFT uses visual question answering so growers can ask plain language questions about tomato leaf disease from photos.
Latest this week
BlockchainOct 22, 20264 min read
Behavior Aware Fraud Checks Proposed for AI Agent Payment Flows
An October 2026 preprint pairs behavior aware fraud scoring with controls on AI agent payment authorization.
BiotechOct 22, 20264 min read
Blitz Preprint Targets Faster Diffusion Steps in AlphaFold3 and Boltz 2
Blitz aims to speed the denoising core shared by AlphaFold3 and Boltz 2 structure pipelines in a 16 October 2026 preprint.
RoboticsOct 22, 20264 min read
EmbodiedSmith Loop Raises Sim Task Success on Robot Demo Benchmark
EmbodiedSmith co evolves sim scenes and robot tasks and reports higher success on hard jobs in an October 2026 preprint.
QuantumOct 22, 20264 min read
Local Measurement Theory Clarifies Readouts on Quantum Cellular Automata
A 19 October 2026 arXiv preprint spells out what local measurements can reveal on quantum cellular automata grids.
AIOct 22, 20262 min read
Sakana Continuous Memory Machine Splits Short and Long AI Memory
Sakana AI posted a two bank memory design on arXiv that may help agents keep fresh notes and older facts without one crowded loop.
LooksmaxxingOct 21, 20264 min read
Face Analysis AI Study Matches Photos to Device Measured Skin Values
Peer reviewed skin research on 1,099 Korean adults shows how face analysis AI tracks visual grades and lab device readings across three camera types.
Cyber SecurityOct 21, 20264 min read
Distributed Multi Agent Attacks Hide Harmful Goals From Monitors
Security researchers reported a frontier monitor caught about 7 percent of split agent attacks versus 87 percent when one agent did the whole job.
Supply ChainOct 21, 20264 min read
Neural Operator Planning Under Noise Targets Uncertain Supply Decisions
An arXiv preprint applies neural operators to planning when transitions are noisy, a pattern inventory and logistics teams simulate often.
AIOct 20, 20263 min read
DeepSeek Open Sources DeepInfer for V4.1 Flash Inference
DeepSeek released its production inference engine as open source with offline paths for V4.1 Flash on NVIDIA SM100 and Ascend 950, giving teams a way to benchmark real serving behavior before they spend on hardware.
AIOct 19, 20264 min read
Anthropic Claude Code on the Web Runs Async Agent Pull Requests
Anthropic added Claude Code in the browser. Cloud agents can edit GitHub repos and open pull requests. Security teams should treat web sessions like CI runners with repo access.
AIOct 18, 20263 min read
OpenAI Adds Ultrafast Tier for Agent Workloads
The OpenAI Python SDK 3.35.0 adds an ultrafast service tier for Agents plus typed Live sideband events, so teams can test latency before they move production tool loops.
QuantumOct 20, 20264 min read
Wannier90 4.0.3 Tightens Library Mode for Materials Teams
Wannier90 version 4.0.3 adds library helpers, pytest based tests, and MPI naming cleanup so quantum chemistry pipelines can embed wannierisation with clearer lifecycle and timing hooks.
QuantumOct 19, 20264 min read
Qiskit IBM Runtime 0.43 Adds Calibration IDs and Mid Circuit Gates
IBM released Qiskit IBM Runtime 0.43 with calibration ids and mid circuit gates. Quantum teams should pin versions before they publish hardware benchmarks.
QuantumOct 18, 20263 min read
IBM ibm_berlin Opens EU Early Access on 120 Qubit Nighthawk
Premium and Flex users can now run jobs on ibm_berlin, a 120 qubit Nighthawk QPU in the eu de region with faster two qubit gates than ibm_miami, though IBM still lists exploratory limits on repetition time and dynamic circuits.
RoboticsOct 20, 20264 min read
FastVLA Runs One Vision Language Call Per Robot Episode
FastVLA keeps a frozen 4B vision language model as task memory and uses a lightweight actor each step, cutting per step inference cost on RoboTwin and LIBERO benchmarks according to its October 2026 code release.
RoboticsOct 19, 20263 min read
PhysHSI Open Stack Targets Unitree G1 Humanoid Scene Tasks
InternRobotics shared PhysHSI tasks for Unitree G1 robots. Buyers should test sim checkpoints before they pay for field trials with heavy contact.
RoboticsOct 18, 20264 min read
SenseNova RoboRSI Publishes LIBERO Code Before 27B Preview Weights
OpenSenseNova released SenseNova RoboRSI code on 16 October 2026 for LIBERO style benchmarks, with open 27B preview weights slated for 20 October 2026 so teams can test embodied agents in sim and on hardware.
BiotechOct 20, 20264 min read
OpenDDE v1.2.0 Adds Constraint Guided Antibody Design
Aureka OpenDDE v1.2.0 introduces TFG constraint guidance so teams can steer antibody antigen co folding with partial interface knowledge without retraining weights.
BiotechOct 19, 20263 min read
GenMol V2 Extends SAFE Syntax for NVIDIA Drug Discovery Models
NVIDIA BioNeMo released GenMol V2 with new SAFE text rules for drug like molecules. Chemistry teams should rerun hit lists before 2026 lab plans are set.
BiotechOct 18, 20263 min read
NVIDIA BioNeMo Drug Discovery Pipeline Skill Joins the Agent Toolkit
NVIDIA added a drug discovery pipeline skill to the BioNeMo Agent Toolkit and merged catalog work in NVIDIA skills pull 654 on 1 October 2026 so coding agents can chain GenMol, DiffDock, and Boltz2 NIM steps.
BlockchainOct 20, 20263 min read
Ethereum Heze Devnet Tests Ship FOCIL and Quick Slots Fixtures
Ethereum execution specs shipped version 2.0.0 of the heze devnet test bundle with conformance cases for EIP 7805 FOCIL and EIP 8198 quick slots on the Bogota devnet fork ahead of broader client work.
BlockchainOct 19, 20263 min read
RISC Zero zkVM Guest Memory Flaw Threatens ZK Guest Programs
RISC Zero patched a guest memory flaw in sys_read inside zkVM. Teams proving blockchain workloads must rebuild guest binaries on fixed SDK releases.
BlockchainOct 18, 20263 min read
OWASP Smart Contract Top 10 2026 Maps SC01 Through SC10 Risk Themes
OWASP published a forward looking Smart Contract Top 10 for 2026 with categories SC01 through SC10 so audit and protocol teams can align reviews with recent incident patterns rather than legacy DeFi checklists alone.
Agri TechOct 20, 20264 min read
RL SMPC Blends RL and Stochastic MPC for Greenhouse Lettuce
A Control Engineering Practice paper documents RL SMPC, which merges reinforcement learning policies with scenario based stochastic MPC to control lettuce greenhouses under parametric weather uncertainty.
Agri TechOct 19, 20263 min read
Infrastructure Free UAV Rice Dataset Opens Low Cost Agri Vision
Researchers published a DJI rice UAV dataset without ground control points. Agri tech teams can reuse the open pipeline for low cost field vision tests.
Agri TechOct 18, 20263 min read
Microsoft FarmVibes AI Hardens CI Pins and Remote AKS API Access
Microsoft FarmVibes AI merged pull 253 on 9 September 2026 to pin GitHub Actions to full SHAs and pull 254 to stop exposing the remote AKS API on the public internet, tightening the ag observability stack.
Supply ChainOct 19, 20263 min read
Sam Altman Bets OpenAI Custom Silicon Stays on TSMC Flow
Sam Altman said OpenAI wants TSMC to add capacity instead of moving custom chips to Intel. Supply chain leads should stress test single foundry plans for 2026 AI hardware.
Supply ChainOct 18, 20264 min read
Broadcom AI Semi Revenue at 10.8 Billion Tests CoWoS Capacity Binds
Broadcom reported about 10.8 billion dollars of AI semiconductor revenue in Q2 fiscal 2026 and guided Q3 near 16 billion while TSMC said CoWoS advanced packaging stays booked through 2026, a tight link for custom ASIC supply plans.
Supply ChainOct 17, 20262 min read
NVIDIA Cosmos Curator 2.9.0 Rewrites Multi Camera Robot Data
NVIDIA shipped Cosmos Curator v2.9.0 on 16 October 2026 with multi camera embeds and robot action video parsing. Supply chain ML ops teams must rebuild manifests before retraining route agents.
Cyber SecurityOct 19, 20263 min read
Cursor CLI MCP OAuth Flaw Enabled Agent Command Injection
Cursor patched a CLI flaw where untrusted MCP OAuth servers could inject shell commands. AI platform teams must vet MCP servers like CI plugins.
Cyber SecurityOct 18, 20263 min read
Hyperledger Besu Patches QBFT DoS and EIP 7702 Consensus Splits
Besu maintainers published GitHub security advisories between 7 and 14 October 2026 covering QBFT denial of service, EIP 7702 consensus splits, and RLP wrapped typed transaction mismatches that could fork private networks.
Cyber SecurityOct 17, 20262 min read
CISA Orders Fast Patch for BeyondTrust Remote Support Flaw
CISA listed BeyondTrust Remote Support CVE 2026 102764 on 16 October 2026 with a 19 October mitigation due date. IT teams should apply BT26 05 guidance on any support appliance.
LooksmaxxingOct 19, 20264 min read
Face Analysis AI Paper LMOL Ranks Photos With Pairwise Tests
Researchers shared LMOL face analysis AI that ranks photo pairs on SCUT FBP5500. App teams should treat beauty scores as fun ranks, not medical facts.
LooksmaxxingOct 18, 20264 min read
Face Attractiveness Test Study Shows Image Reward Models Favor Pretty Faces
The Biased Attractiveness Bench paper finds image reward models often treat a face attractiveness test score like a quality score, which can skew diffusion tuning and mislead anyone using AI face ratings for fun or research.
LooksmaxxingOct 17, 20262 min read
Mapika decider 2b vision Offers Calibrated Photo Scores
Mapika released decider 2b vision for one pass photo decisions with calibrated probabilities. Beauty app teams should read model limits before changing face score UI copy.
