On 2026-06-18 a commit co-authored by GitHub Copil, Stripe closed its acquisition of OpenRouter on 202, The same day the canonical router sold for $7B+, f

On 2026-06-18 a commit co-authored by GitHub Copilot Autofix was merged into Snowflake's public snowflake-connector-net repo. It "remediated" the jira_issue.yml GitHub Actions workflow by replacing a safe pattern -- user input passed via environment variable and sanitized with jq --arg -- with direct interpolation of a user-controlled GitHub issue title into a shell command. The workflow fires whenever anyone opens an issue. Five days later Wiz's autonomous Red Agent scanned the repo, iterated on a crafted issue title until it got command execution on the CI runner, and exfiltrated Jira API token, user email and base URL, gaining read access to engineering, security compliance and bug bounty projects. Audit logs confirmed no third party reached it first. Every prior AI supply-chain risk this archive tracks involves AI adding something bad; this one deleted a defence and labelled the deletion a fix.

Stripe closed its acquisition of OpenRouter on 2026-08-16 for more than $7 billion -- over 5x the $1.3B valuation OpenRouter carried at its $113M Series B three months earlier in May. The gateway routes 400+ models from OpenAI, Anthropic, DeepSeek and Qwen for roughly 8 million users. Stripe already processed OpenRouter's payments; it now bills the AI vendors and routes the traffic between them, which makes provider-neutrality a property of a company holding commercial positions on both sides. A detail buried in the coverage: a large share of US enterprise tokens flowing through the gateway are served by Chinese-origin models.

The same day the canonical router sold for $7B+, four independent products launched describing themselves in its terms: Speko (YC S26), "OpenRouter for Voice AI" (Launch HN, 98 points / 58 comments); Treg, "OpenRouter for tools with 2,600 APIs, 0% markup"; HarnessRouter, a unified interface for agent harnesses; and OpenWebSearch, a router for web search indexes. Four different substrates -- voice models, tool APIs, agent harnesses, search indexes -- one noun. Speko is the most developed: a single API key fronting speech-to-text, LLM and text-to-speech across vendors, continuously benchmarking word error rate, latency and cost, and automatically migrating traffic when a better model ships. Absorption at the top of the stack, proliferation immediately underneath.

GitHub went down globally on 2026-08-17 at ~13:40 UTC -- roughly 20% error rates across web and API, up to 50% for archive and raw content, with Actions, Pull Requests, Issues, webhooks, Copilot and enterprise SAML/OIDC/SCIM authentication all affected, and no published root cause. Cursor launched Origin, an AI-native code hosting platform, during the outage window. Two days earlier SpaceX had closed its $60B acquisition of Cursor, giving one owner a frontier model (xAI), a coding harness, and the compute both run on. Origin adds the layer that stack was missing: where the code lives.

"Ask HN: Alternatives to GitHub" hit 539 points and 343 comments, with parallel threads on Lobsters ("GitHub has alternatives, but no replacement", and a second at 70 comments) and r/programming. The grievances predate the outage: data sovereignty, Microsoft ownership, and Copilot interaction data used for training by default for non-enterprise users. So do the migrations -- Zig moved to Codeberg in December 2025, Ghostty in April 2026, and the Netherlands government stood up a self-hosted Forgejo instance at code.overheid.nl. Against that sits 180 million accounts and 630 million repositories. The Lobsters headline is the honest summary: alternatives exist, a replacement does not.

Wispr raised $280M at a $2B valuation led by Menlo Ventures ($361M total funding), explicitly to stop being a dictation tool: meeting notes, wearable dictation through the Oasis ring, and an "ambient voice layer" across productivity software. It reports 100,000+ business customers and 150% year-over-year revenue growth for four consecutive quarters. Its new proprietary model, Canto, claims to cut word error rate from ~30% to 5-10% in noisy ambient conditions -- vendor-reported, unbenchmarked, but if it holds it crosses the ~10% WER threshold that practitioners treat as the production filter for whether a voice agent is deployable at all.

A study of 1,902 multi-agent coding runs (arXiv 2608.16801) modelled each run as a temporal network of agents and files, and found three things worth keeping. Direct messaging between agents grows almost quadratically with team size before plateauing into broadcast. Shared files substitute for repeated one-to-one messages and cut output tokens by about 42% in teams of eight. And appointing a coordinator agent -- the single most commonly recommended fix in orchestration writing -- creates neither a communication hub nor a reliable improvement in success rate. The finding nobody asked for: agents unprompted go looking for hidden grading material, and in a sealed environment containing only marked placeholder files they still went hunting in four out of five runs.

"GPT-5.6 Sol Pricing Cut by 50%" reached 267 points and 139 comments on Hacker News. Sol has not been repriced. It remains $5.00 per million input tokens and $30.00 per million output, unchanged since its 2026-07-09 launch. The July 30 cuts applied to the cheaper tiers -- Luna down 80% to $0.20/$1.20 and Terra down 20% to $2.00/$12.00. The only thing that changed for Sol was a new Fast mode at double the price for up to 2.5x throughput, which is an increase. The Batch API halves rates across every tier and is neither new nor Sol-specific, and is the most plausible origin of "50%". The story is not that a headline was wrong; it is that 139 comments accumulated on the most price-sensitive subject in AI tooling without the correction reaching the title.