Newsdesk
Engineering
Data & AI
Industries
Enterprise Systems
Go-to-Market
Longform
DesignIndiaAll stories

AI Tech Weekly Digest

13 November 2026

68 new posts across 9 AI lab and practitioner sources — 7 worth your time.

One week of AI Tech Weekly Digest, 7 stories, as published.

Security

Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents

Microsoft researchers detailed 'Lazarus,' a backdoor technique that reprograms dormant neurons in AI models to execute hidden malicious commands upon detecting specific triggers. The vulnerability is severe for autonomous agents equipped with system terminal access or tool-calling capabilities, as triggered models can silently run remote code or exfiltrate secrets.

Why it matters — Autonomous agents given terminal or tool execution access require secondary execution sandboxing beyond prompt safety guards to mitigate latent neural backdoor triggers.

Microsoft Research · 13 November 2026 · Read the original →

Research

Introducing AlphaProof Code: a breakthrough in verifying complex software at depth

Google DeepMind introduced AlphaProof Code, an agent system that generated three million lines of formal proofs in Lean. The system was applied to formally verify the core encryption implementation of the Signal protocol.

Why it matters — Formal verification of security-critical codebases can be automated at scale by coupling code agents with formal proof assistants like Lean.

Google DeepMind · 13 November 2026 · Read the original →

Tools

Quoting Prompting Claude Fable 5.2

Anthropic's testing of Claude Fable 5.2 showed that code modification behavior changes significantly based on prompt ownership framing. When a prompt presents code as belonging to someone else, the model appends duplicate logic, but when framed as the model's own codebase, it aggressively consolidates and refactors code.

Why it matters — Explicitly declaring codebase ownership as the model's own in prompt instructions encourages consolidation over additive code duplication.

Simon Willison · 13 November 2026 · Read the original →

Models

Gemini Decisions API

Google launched the Gemini Decision 1 API, offering structured classification endpoints for yes/no, multiple choice, and rubric-based scoring. The service is priced at $0.10 per million input tokens with free output tokens, matching OpenAI's GPT-6 Luna decision pricing.

Why it matters — Offloading simple classification and rubric evaluations to dedicated decision APIs reduces execution costs compared to general-purpose LLM completion calls.

Simon Willison · 13 November 2026 · Read the original →

Tools

NVIDIA Dynamo Recipes Bring Korea’s Leading Open Models to Production

NVIDIA released Dynamo deployment recipes designed to run large open foundation models across multi-GPU nodes. The recipes handle request routing and inter-GPU communication to maintain throughput when models exceed single-GPU VRAM limits.

Why it matters — Deploying open foundation models that exceed single-card VRAM capacity requires dedicated request routing orchestration to prevent communication overhead from bottlenecking inference throughput.

NVIDIA · 13 November 2026 · Read the original →

Hardware

Nvidia privately confirmed it could track chips to fight smuggling, even as industry publicly questioned feasibility

Nvidia representatives privately confirmed to U.S. lawmakers that technical mechanisms for tracking AI chips to prevent illegal smuggling under the Chip Security Act are feasible. This acknowledgment directly contradicts public statements from industry trade groups asserting such tracking is technically unworkable.

Why it matters — Hardware-level tracking for advanced AI accelerators is technically viable, paving the way for enforceable semiconductor supply chain tracking mandates.

MIT Tech Review · 13 November 2026 · Read the original →

Business

Kuken Industries searches 70 years of drawings in minutes

Kuken Industries deployed the OpenAI API to search seven decades of legacy cooling tower engineering schematics. The deployment reduced repair document retrieval time for engineers from nearly an hour to 4.5 minutes.

Why it matters — Indexing complex legacy visual schematics through multimodal APIs yields a 90% reduction in technician retrieval workflow time.

OpenAI · 13 November 2026 · Read the original →

Defending dormant neurons before lunch, searching seventy-year-old drawings after.

7 stories, every Friday

Published here every week. Follow by RSS to get it as it lands.

← Previous issue Next issue →