AI-found Linux kernel RCE, one token budget for ChatGPT and Codex, and DeepSeek's open-source repo agent

Aviator Verify — a seven-part review recipe from Aviator's CEO: write your team's recurring review comments down as "invariants" — standing rules an agent checks on every change — capture intent and decisions from the coding session, then let humans judge evidence instead of reading every line. It targets the complaint your reviewers already have: the diff never shows what the agent decided, so review moves to intent and proof. The suggested ramp is three months, with reviews shrinking as the rule registry fills. Invariants are drafted from your team's existing PR history and go live only once the owning team approves them. The New Stack

“Judgment needs intent, and accountability needs a person.”

dsh-superagent — a free, MIT-licensed research prototype of an agent that lives on one GitHub repository: approved people ask for work, it opens a work item and answers with a pull request. The README warns it is research code, not an official DeepSeek product. Worth a skim before you write your own autonomy policy: its defaults are deliberately cautious — a dedicated GitHub account, operator-only access at first, an observe mode that logs what it would have done, and auto-merge off. 7.1k stars on day one; the team's design essay argues one person can track only three to five concurrent agent sessions. GitHub

OpenAI allowance merge — From December, Plus, Pro and Business plans get a single token-based allowance shared across ChatGPT and Codex, as the Work agent features move into the main app; Pro 100, 200 and 500 allowances were doubled on October 29 after backlash over a planned cut. Seat maths change: a heavy coding day now eats the same budget as research and document work. OpenAI has not published per-plan token figures, so ask for them before a December renewal. Quick Instant chat does not count towards the shared pool; combined usage shows under Settings > Usage. BleepingComputer

Hacker News thread — A 448-point thread argues over a claim, which Anthropic has not confirmed, that the company bans users whose chats touch training larger models. The advice that keeps recurring is that a team should not build a critical path on one vendor: keep a working alternative and portable workflows. Others counter that model churn makes accumulated technique worthless anyway, so use the best model and switch at the guardrail. The practical manager version: know which of your workflows could actually move this quarter. Hacker News

Kyle Kingsbury — Kyle Kingsbury, whose Jepsen tests have audited distributed databases for over a decade, writes that LLM use erodes trust in documentation and software because the signals engineers relied on to judge a mature project — thorough docs, polished releases — are easier to counterfeit. For vendor picks and dependency reviews, "looks professional" is no longer evidence; ask what a project can prove rather than what it can present. The quote comes from Jepsen's results for ParticleDB 1.1.4, relayed by Simon Willison. Simon Willison's weblog

“the signals we used to rely on as markers of a mature system are now easier to counterfeit.”

VAPOR PANDA — CrowdStrike details an espionage campaign that aimed AI agents at 200+ organisations across 14 countries — intrusions now scale with agent frameworks, one consuming 1.4 billion input tokens. CrowdStrike

Tampered Claude Code extension never reached official registries. Microsoft confirms the tampered Claude Code build never reached official registries; how it entered Cursor's marketplace is unknown, so keep extension auto-update policies in place. Microsoft Security Blog