Agent-built exploits hit production kernels; OpenAI merges Chat and Codex allowances; DeepSeek open-sources a team coding agent

HackerTube — a YouTube-style site Karpathy built in about two hours with an agent: it collects 6,454 videos recommended in Hacker News comments since 2007, sorts them across 150 topics and serves its own recommendations with no user accounts. A personal project, not a product. A crawl-and-rank consumer app is now a one-person evening build, so "a competitor could rebuild us in a weekend" is a roadmap question you have to answer, not a joke. The Hacker News thread drew 2,162 points, with commenters surprised how closely an unpersonalised feed matched their taste. karpathy.bearblog.dev

dsh-superagent — an open-sourced research prototype — explicitly not a product, and largely written by coding agents — of a proactive agent for one GitHub repo: an approved person asks for work, it opens a work item, runs the job in its own session and replies with a pull request. A published reference for giving an agent a seat in a team's workflow; its defaults (a dedicated GitHub account, observe mode, auto-merge off) are a ready checklist for anyone scoping something similar. About 7.1k stars in its first day; the README warns it is research code that may still contain bugs. GitHub

Armando Picón — a practitioner's working position on agent-assisted delivery: agents take the implementation, engineers keep the accept-or-reject call. Durable decisions live in a small reviewable home (his example: a PROJECT_GUIDE.md), and every increment gets explicit acceptance cases written before any code exists. Acceptance cases are the spec format agents can actually act on — a cheap thing to try on your next ticket before betting the roadmap on spec-driven tooling. picon.dev

Jaclyn Konzelmann — a paid podcast with Google Labs' product director on how Gemini 4 changed her job: with code effectively free, deciding what deserves to exist is the work, and she bans AI-written docs — her PMs must be able to defend every statement themselves. A counterweight from inside a big platform: generation is cheap, judgment is the scarce input, and document quality is where she polices it. Only the takeaways are public; the full episode sits behind Behind the Craft's paywall. creatoreconomy.so

“It shouldn't take someone longer to read your doc than it took for you to write it.”

The AI cost of JavaScript frameworks — Nolan Lawson had an agent build the same small Pokémon app in twelve JavaScript frameworks, three to five runs each, and priced the runs: $2.67 median on Hono+htmx versus $4.97 on Angular, which twice failed to produce a working app at all. Framework choice now carries a token-cost line, and cost-per-agent-build is a number a PM can put in an estimate instead of vibes. His own caveat: one small content-heavy app, and agent runs are stochastic. nolanlawson.com

ChatGPT and Codex move to one shared allowance. From December, Plus, Pro and Business plans draw ChatGPT and Codex usage from a single token-counted allowance as OpenAI folds its agentic Work features into the main app; quick Instant chat stays outside the pool. Pro Work and Codex allowances were doubled on October 29. The clearest signal yet that consumer AI pricing is converging on metered compute — useful calibration if you are pricing your own agent features per seat. No token amounts per plan have been published yet, so buyers cannot compare the pool with today's separate limits. BleepingComputer

A Brazilian court fines lawyers for prompt injection in a filing. A Brazilian labor court censured and fined two law firms after they hid white-on-white text in a filing instructing Galileu, the judiciary's AI review tool, to answer superficially and not challenge the documents; the system flagged and blocked the content automatically. Reported as the first ruling of its kind anywhere — and a direct warning for any product whose agents read files other people control: ingestion is attack surface, and courts now treat it that way. The fine was 10% of an R$842,000 claim, about $16,000. jetstream.security

Proton — After a 404 Media investigation, Andy Yen cancelled near-company-wide interviews in which employees would have had to justify how much AI they use, with jobs potentially at stake; he apologised on Reddit and promised to publish Proton's internal AI policies. The cautionary case for pushing agent adoption by mandate: the same reporting had him calling a security engineer's pentest request "loser thinking" — the costs landed on talent and user trust, not on the tooling. A reorganisation is planned for next month, with no job losses from AI, he said. 404 Media

Sweep — ServiceNow confirmed it acquired the Israeli startup Sweep to speed customers' moves onto its CRM; the price is undisclosed but estimated in the hundreds of millions. Agent-assisted migration is now acquisition-worthy — if you sell enterprise software, the implementation timeline is quietly becoming the product. CEO Bill McDermott says large-company rollouts now take 30 to 60 days. Techmeme (CTech)

Pipecat — Voice AI provider Deepgram acquired Daily, the company behind the open-source voice-agent framework Pipecat, which had raised more than $62M. Voice-agent infrastructure is consolidating fast — pick stack partners expecting them to get absorbed, and check licence terms while the framework is still open. The deal price was not disclosed. Techmeme (TechCrunch)

VAPOR PANDA — CrowdStrike reports a China-linked espionage crew pointed the ARTEX agent framework, running DeepSeek, at more than 200 organisations across 14 countries in 11 days, with breaches confirmed at nine, including ministries and US pharma and semiconductor firms. The agents were explicitly told to find machines without the CrowdStrike sensor — assume automated reconnaissance at scale in your threat model. Planning agents burned through about 1.4 billion input tokens on one intrusion alone. CrowdStrike