What a read-only Postgres role does and doesn't stop when an AI agent connects
A reproducible PostgreSQL 17.11 run shows which harms a read-only role blocks and which session settings and reads still get through.
AI Daily Recap
44 articles · 6 categories
The finishable daily brief
What happened in AI — Nov 10, 2026
Tuesday, Nov 10, 2026
44 articles · 6 categories
read top to bottom · then stop
In 30 seconds
- Gemini 4 Carbon is generally available, with a new multimodal agent benchmark (MKA) alongside it.
- DeepMind's AlphaProof Code formally verified security properties of Signal's real protocol code.
- NVIDIA is donating its OpenShell agent-governance runtime to the CNCF.
- A read-only Postgres role does not stop every harmful agent query; a reproducible test shows which ones.
- Fireworks AI's breach now includes billing data and some customer secrets; Notion is notifying users.
- Epoch: GB300 has 25x H100's agentic inference capacity on only 2.5x the FLOP/s.
Agent safety moved to enforcement points: tools that intercept commands, audit access and test database roles shipped alongside NVIDIA donating OpenShell to the CNCF.
Google made Gemini 4 Carbon generally available and DeepMind formally verified Signal's protocol code with agents, while the Fireworks AI breach spread to customer data.
Prefer to watch?
LLM Digest Daily Explainer: Nov 10, 2026
Guardrails on What Agents Can Touch 5 items
Builders are moving agent safety from prompts to enforcement points: database roles, command interceptors, access audits and a CNCF-bound agent runtime.
Antelog – stops AI coding agents from running dangerous commands
Local interceptor that checks each shell command a coding agent proposes and pauses destructive ones, like deleting a production volume, for approval.
I Access Auditor – see and revoke what your AI coding agents can access
Desktop app that inventories what Claude Code, Codex, Gemini CLI, Cursor and OpenCode can access, and lets you revoke it.
NVIDIA Expands Open Source Commitment With OpenShell Donation and New AI Infrastructure
NVIDIA is donating OpenShell, its runtime for securing and governing autonomous agents, to the CNCF as a Sandbox project.
Show HN: Fix Gate – verifies security-fix PRs, incl. AI-written ones
Rule-based GitHub Action that checks a security-fix PR: finding gone, nothing new, files parse, tests not newly failed; unrun checks never count as passes.
Agent Experience: Docs and APIs Now Have an Agent Audience 4 items
Agent traffic to docs is now large enough to measure, and vendors are starting to design SDKs and docs for the agent that picks them.
Measure agent-traffic for any site
Open-source tool from Mintlify to measure how much of a site's traffic is agents, which ones, and whether a GPTBot request is really OpenAI.
What Is Agent Experience (Ax)?
Microsoft defines AX: an agent picking a rival service or calling a deprecated API makes your product look broken, so measure it.
Designing for Agents — Zeno Rocha, Resend
Resend's Zeno Rocha on where answer-engine optimization ends and agent experience begins, from docs Claude Code reads to tools it calls.
Show HN: Get your coding agent off ~2019 JavaScript, onto modern web primitives
Agent skill that steers coding agents off legacy JavaScript patterns toward native browser primitives like CSS.highlights and dialog.
Verifying and Operating Agent-Built Software 4 items
The verification bar for agent output is rising, from formal proofs of real protocol code to evidence that messy codebases derail coding agents.
Introducing AlphaProof Code: a breakthrough in verifying complex software at depth
DeepMind's agentic framework formally verified key security properties of the actual Signal protocol code.
Google Finds that Technical Debt Can Derail Coding Agents Through "Agent Distraction"
Google research finds technical debt causes "agent distraction" that throws coding agents off task.
Article: LLM Reliability Layers Under Test: Accuracy, Yield, and the Limits of Verification
InfoQ tests LLM reliability layers on accuracy and yield, and maps where verification stops helping.
How Workday Built an Enterprise-Scale AI Agent Platform with LangGraph and LangSmith
Workday's internal agent platform uses LangGraph for orchestration and LangSmith for tracing and evals, targeting production agents by end of 2026.
Frontier and Small Model Releases 5 items
Three labs shipped on the same news cycle: Google's top tier, Anthropic's Fable 5.2, and Meta's smallest Muse model.
Gemini 4 Carbon is now available
Google's strongest model is now generally available, pitched on cybersecurity and long-horizon agentic work, after a two-week defender-only preview.
The Multimodal Knowledge Agent Benchmark
New Google and Artificial Analysis benchmark for agents that must reason over visually complex files, not just text.
[AINews] Fable 5.2 retakes World #1 LLM spot as $ANT IPO roadshow begins
Latent Space covers Fable 5.2 and Mythos 5.2 retaking the top LLM spot as Anthropic's IPO roadshow opened.
Introducing Our Smallest, Most Efficient Model: Muse Ember
Meta's small, efficient model enters the Haiku/Luna tier, per Simon Willison's notes.
Kimi K3.1 Pro Coder Arrives at 775 GB, Built to Fit One GPU
Moonshot's coding model ships as a 775 GB release aimed at fitting on a single GPU.
Inference Economics: Memory, State and Mixed Hardware 4 items
Inference gains now come from memory, state reuse and plumbing more than raw FLOP/s.
Beyond FLOP/s: comparing AI hardware for inference by memory and communication
Epoch estimates GB300 delivers about 25x H100's agentic inference capacity per chip with only 2.5x the 8-bit FLOP/s.
FlashStates Unlocking Flexible Hybrid State Management for Agentic LLM Inference
Together AI and collaborators decouple recurrent checkpoints from KV-cache blocks so hybrid models can resume agent sessions without dense state storage.
CUDA+ROCm dual-back ends LLM inference for mixed-vendor GPUs (up to 48% faster)
Patch splits GLM MoE inference across an RTX 5090 and a Radeon iGPU, CUDA for attention and HIP for experts, reporting up to 48% faster.
Show HN: QueenMQ – slow LLM calls blocked our Kafka partitions, so we built it
Apache-2.0 message broker built after slow LLM calls blocked Kafka partitions; it has mostly replaced Kafka at its authors' company.
Inference Vendor Breach Widens 2 items
The Fireworks AI incident now reaches customer data, and downstream apps are notifying their own users.
Fireworks.ai Update on October 2026 security incident
Fireworks now reports impact to account and billing data and, for some accounts, secrets, datasets and incomplete models.
Tell HN: Notion AI chat sessions have been exposed via Fireworks AI hack
Notion is emailing users that AI chat sessions may be exposed through its vendor Fireworks AI.
You are caught up for this edition