Blog

Notes on building reliable AI products.

A GPT-Live audio waveform carrying a SynthID trace through a verification gate into a logged evidence receipt

OpenAI Watermarks Supported GPT-Live Audio. The Evidence Trail Is Still Yours

OpenAI adds SynthID to supported GPT-Live audio and opens verification access. Here is the evidence workflow voice-agent builders now need.

DeepSeek V4-Flash-0731 checkpoint flowing through an agent harness into an outcome-based routing gate

DeepSeek V4-Flash-0731 Makes the Harness Part of the Model

DeepSeek V4-Flash-0731 adds major agent gains, Codex-ready Responses API support, and open weights. Here is how builders should evaluate it.

MiniMax H3 multimodal inputs entering an asynchronous render queue and emerging as a finished 2K video asset

MiniMax H3 Gives Media Agents a 2K Render Queue

MiniMax H3 brings 2K video, native audio, mixed references, and async API jobs. Here is the architecture, pricing, and open-weights catch.

LG’s K-EXAONE 2.0 Makes Open Weights a Cluster Procurement Decision

LG’s 750B K-EXAONE 2.0 is Apache 2.0, agent-ready, and built for eight H200s. Here is what its open-weight economics mean for builders.

Inkling-Small open-weight model routing coding, tool, knowledge, and frontier escalation tasks

Inkling-Small Puts Open Weights Back in the Agent Router

Thinking Machines’ Inkling-Small pairs 12B-active MoE compute with open weights. Here is where it fits—and where it should escalate.

A simulated Claude cyber range split open to reveal real internet systems and a mismatch between declared and effective scope

Anthropic’s Claude Cyber Evals Hit Real Organizations. The Simulation Prompt Was Wrong

Anthropic says six Claude cyber-eval runs reached real organizations. The incidents show why prompts, vendor paths, and side effects need hard controls.

Gemini Robotics ER 2 persistent agent session above robot tools and an independent safety rail

Gemini Robotics 2 Makes the Persistent Session the Robot’s Control Plane

Google’s Gemini Robotics 2 stack opens ER 2 as a stateful robot agent while keeping action models gated. Here is what builders must test.

RufRoot cover showing one exposed MCP bridge fanning into shell execution, AI provider keys, conversations, agent swarms, and persistent memory

RufRoot Turned One MCP Port Into an Agent Compromise Domain

CVE-2026-59726 gave unauthenticated attackers RCE, provider keys, conversations, swarms, and durable agent memory writes. Here is the builder response.

Terraform plan changes flowing through review into an OpenAI project governed by explicit identity, model, tool, retention, limit, and drift controls

OpenAI’s Terraform Provider Gives AI Governance a Plan File

OpenAI’s official Terraform provider makes projects, access, model policy, rate limits, and drift reviewable—but its gaps matter just as much.

Codex CLI 0.146.0 architecture showing a portable plugin core, Codex control plane, and remote execution nodes

Codex CLI 0.146.0 Moves the Agent Boundary Out of the Terminal

Codex CLI 0.146.0 adds Agent Plugins and remote Code Mode, turning portability, execution, network policy, and MCP state into runtime concerns.

MCP 2026-07-28 architecture showing visible stateless requests routed through a policy gateway while agent work state remains on a separate rail

MCP 2026-07-28 Deleted the Transport Session. Your Agent Still Needs One.

MCP 2026-07-28 removes protocol sessions, exposes agent calls to gateways, and moves durable state into explicit handles, tasks, and runtimes.

Codex Security reviewing a repository and routing CI to pass, finding failure, or incomplete-evidence failure

OpenAI’s Codex Security CLI Makes Missing Evidence Fail the Build

OpenAI's open-source Codex Security CLI turns findings, coverage, scan history, and remediation into a CI contract builders can audit.