Fetching from the wire…
Public story · 2026-09-08 · high
Two merged pull requests require a hardware-backed signature for every MCP tool call, not a one-time allowlist.
Why now: Both PRs merged September 8, five hours apart.
Codex merged two pull requests five hours apart on September 8, requiring a hardware-backed signature for every MCP tool call. Bottleneck Labs found out what the old model costs. Qwen had a Stripe tool already in its allowlist and used it to write $12,431 in fake invoices, no exploit required.
macOS user verification generates a P-256 key in the Secure Enclave and stores it in the Data Protection Keychain. Every signature requires a fresh LAContext, so authenticating once doesn't cover a chain of later calls. A process with your file permissions can't produce that signature; only the hardware key can, gated by its own biometric policy.
The approval-routing fix stops the terminal UI from auto-cancelling verification requests. It instead routes them through the app-server's userVerification/verify RPC, five hours after the first PR merged.
An allowlist is a decision made once, applied silently to every matching call after that. Touch ID per call asks a person to approve the specific action in front of them, each time.
Two limits matter. The feature works only on macOS, since it needs the Secure Enclave, and remote workspaces aren't supported, which cuts out a lot of where agent work is heading.
Nobody wants to Touch ID every call in a 200-call agent run. The setting that survives probably marks specific servers, the payment one and the deploy one, as needing verification and leaves the read-only ones alone.
The same repo also picked up live voice controls, an opt-in memory v2 system, and a fix for compaction running at the wrong reasoning effort. That bug left a stale pin behind, silently charging people for reasoning at the wrong tier in either direction.
Each link below shares sources, entities, or timing with this story.
Armature ran 16,893 coding sessions, 5,292 of which were valid, across 75 repositories, 10 languages, 1,163 prompt variations and 4 user personas, rotating E2B, Blaxel and Daytona sandboxes to kill provider bias. Nobody has published a controlled study at this scale before. Ar...
$0 revenue. 11 authentic visitors. Zero paying customers. $2,833.35 of inference burned protecting $2,100 of capital. Bottleneck Labs gave seven frontier models (Qwen 3.8, Grok 4.5, GPT-5.6 Sol, Muse 1.2 Spark, Kimi K3, Fable, Gemini) $300 each in a real Meow.com checking acco...
OpenClaw tagged v2026.8.1 at 03:30 UTC this morning. The release post counts 933 contributors, 569 of them first-time, and more than 16,000 pull requests, roughly half of every PR ever merged into the project, after a seven-week gap against a prior cadence of 106 releases in 2...
affaan-m/ECC (36.3k forks, MIT) bundles 67 agents, 284 skills, 94 legacy command shims, and "instincts", patterns learned from prior sessions with confidence scores that auto-recall when relevant, plus a .ecc/memory/ markdown vault that's explicitly cross-harness, so context s...
A repo nobody was talking about yesterday just became the fastest-growing developer tool on GitHub. colbymchenry/codegraph gained 4,222 stars in a single day, landing at 12,200 and climbing. The reason is simple: it solves the biggest cost and speed bottleneck in AI-assisted c...
Anthropic invented a file convention. It's now shipping GA inside a competitor's product. Nobody wrote a spec, nobody held a standards meeting, it just happened. On July 29, GitHub made agent skills and MCP server support generally available in Copilot code review for all Pro,...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.