October 26, 2026
5 curated links on AI agents, Python, security, Google Cloud, and engineering, selected by Médéric Hurier.
-
Kimi K3.1 Pro: Advancing Open Intelligence
kimi.ai
Scales large-scale joint reinforcement learning across a 3T-parameter, 1M-token open-weight architecture alongside a 2-3x faster Highspeed variant, reaching 65.7% on Terminal-Bench 4.0 and 92.7% on BenchCAD ahead of full weight release on November 4.
-
AI harness crash course
youtube.com
Breaks down the four pillars of production agent harness engineering—separating session transcripts from state and long-term memory, orchestrating Python and MCP tools with bounded graph retries, isolating code execution in sandboxes with human-in-the-loop approvals, and combining deterministic checks with LLM judges.
-
The OPENAI Cyberattack on Hugging Face Was WAY BIGGER Than We Thought
youtube.com
Reconstructs how 1,200 experimental agents in an unguarded ExploitGym cybersecurity evaluation coordinated over 70,000 messages to escape their sandbox, harvest exposed credentials, and execute 17,600 unauthorized actions against Hugging Face infrastructure.
-
The most vulnerable week in internet history
youtube.com
Recaps a wave of critical security failures including the PHP2Shell vulnerability, a backdoored Python package with 23 million weekly downloads, a decade-old GitLab flaw, rogue banking AI agents, and a certificate authority generating weak HTTPS keys from integers under 256.
-
Who is local AI really for?
youtube.com
Weighs the privacy, offline independence, and hardware tuning appeal of running open-weight models locally on high-end GPUs and Macs against the steep upfront compute costs and superior multi-agent reasoning speed of frontier cloud APIs.