Posts

Anthropic disclosed in September 2026 that a weapons-engineering cell in Houthi-held northern Yemen used its coding agent Claude Code to develop guidance, navigation, and control (GNC) software for three simultaneous missile projects over eight months. The case is the most concrete example yet that commercial AI coding tools are a genuine dual-use technology—capable of legitimate software work and, in the wrong hands, serious harm. Anthropic found no fielded operational weapon, but the details force governments and vendors to confront how quickly non-state actors can now attempt state-level precision munitions engineering. ...

Every Claude Code Web session you run doesn’t execute on a shared server — it boots inside a dedicated Firecracker microVM managed by an undocumented Anthropic platform codenamed Antspace. That is the finding that emerged in March 2026 when developer AprilNEA ran a single strace -p 1 command inside a Claude Code Web sandbox and exposed a five-layer architecture: a Firecracker microVM, an unstripped 27MB Go runner binary, a Rust process supervisor, a mature but completely undocumented deployment PaaS, and a web-app builder codenamed Baku. This post walks through how the discovery was made, what each layer reveals, and why it changes how you should think about isolating your own agent workloads. ...

Autoprompt is an open-source skill for coding agents that wraps planning, implementation, testing, review, and repair around tools like Claude Code, Codex, and OpenCode. According to the project’s own benchmark run, it solved 73 of 89 Terminal-Bench 2.1 tasks versus a 60-task baseline, raising pass rate from 67.42% to 82.02% and cutting failures by 45%. The trade-off is roughly 3x execution time and 2x tokens, making it most valuable on hard, ambiguous tasks. ...

Benjamin-Plus is a JetBrains-published, MIT-licensed token-efficiency skill for coding agents that measurably cuts cost per task by up to 18% and total tokens by up to 22% without detectable quality loss. In this 2026 review we walk through the independent paired A/B results, the five habits it teaches, the decisive finding that you must inject it rather than install it, and a practical adoption checklist for Claude Code, Codex, and Cursor. ...

OpenBot is CopilotKit’s open-source, MIT-licensed platform where every AI coworker gets a computer of its own — a real browser with its own logins, its own files, and only the tools you grant. It runs self-hosted on Docker Compose with your own PostgreSQL, accepts any AG-UI agent, and gates every action through a fail-closed policy before it happens. As of late September 2026 it sits at roughly 5,200 GitHub stars, 663 forks, and 29 contributors. ...

only-cli (oc) 2026 Guide: Turn Any Website into a CLI Tailored for AI Agents Agent tools waste most of their token budget on raw HTML markup that a model can never use. only-cli (oc) solves this by distilling any website into a compact, numbered CLI view that AI agents can drive by number, cutting token use by about 45x to 118x versus raw HTML. This guide covers what oc is, how to install it, wire it into Claude Code, Codex, Cursor, and Copilot, and how to handle logins and bot blocks safely. ...

Coding agents call MCP servers from scripts by running a CLI (like mcp-call), writing a one-shot script against the MCP SDK over stdio, or letting Code Mode generate code that talks to an HTTP/streamable MCP server. In 2026 you register the server once with claude mcp add or codex mcp add, then your agent’s bash or Python script invokes tools, pipes results through jq, and orchestrates several MCP tools in a single shell flow — instead of burning token budget on dozens of native tool calls. ...

Headless IDA Pro MCP lets AI agents analyze binaries directly through IDA’s decompiler and disassembler without opening the GUI. You set up IDA Pro 9.4, expose idalib through a Rust-native MCP server such as ida-mcp-rs, connect it to Claude Code or Cursor, and then drive decompilation, xrefs, renames, comments, and scripting through natural language. In 2026 this workflow is faster to stand up, more memory-safe, and far more capable than the Python MCP servers of a year earlier. ...

Electron’s safeStorage is not strong enough to protect the API keys your AI agent app stores. To secure OpenAI, Anthropic, or JWT credentials against other processes and malicious code, use the macOS Data Protection Keychain directly with kSecUseDataProtectionKeychain: true, restrict access to a code-signing access group, and gate decryption behind Touch ID or a password. This guide walks through the threat model, the code-signing setup, and a working implementation that keeps a dozen backgrounded agents from reading each other’s secrets. ...

The dsh trace compare agent exploration use case is the fastest way to see why a DeepSeek Harness agent does what it does: install the dsh-maze plugin (formerly dsh-trace-compare), load a session log or follow a live run, and you get the agent’s full exploration maze — main path, failed detours, and backtracking points — rendered on a single timeline with deterministic, LLM-free analysis you can drill back to the raw commands and outputs. It turns an invisible black-box trajectory into something you can actually read, compare across models, and fix. ...