Last Week, Rewired: Examiners, Exploits, and Receipts

Last Week, Rewired: Examiners, Exploits, and Receipts

Welcome to Last Week, Rewired — everything important that happened in AI last week, minus the 400 open tabs. Here's the week that was, November 2–8.

1. The examiners got a lens — and it looks at consequences, not cleverness

On November 4, the OCC issued Bulletin 2026-54, a 20-page framework that tells its examiners how to look at a bank's use of generative and agentic AI. Sixteen components sit under six pillars — purpose, risk, control, oversight, resilience, assurance — with an illustrative four-tier risk classification. An autonomous payment agent sits at the top tier; an internal productivity tool sits at the bottom. The more consequential and autonomous the use case, the more control, human oversight, independent validation, and resilience an examiner will expect to find.

Two things worth knowing: the framework sets no new supervisory expectations and no enforceable standards — it is a resource for examiners, not a rule, and banks are not expected to use it to grade themselves. And the name is not the story. The OCC now calls AI "Super Intelligence," following an October 2 executive order — a rebrand, not a substance change. The substance is the lens.

Mary's take: this is the week the examiner's question got written down in plain sight. Not "how smart is the model?" but "what happens if it gets it wrong, and who answers for it?" That is the same question I keep telling you to ask your vendors — and now there is a 20-page version of it, with tiers attached. If you run AI in a bank, read the framework the way an examiner would: start with your highest-consequence use case and work down.

2. AI found the flaws — 81 in Firefox, 948 working exploits in Linux

On November 5, OpenAI released GPT-6 Cyber, a model built for security research, and rated it "Critical" under its own Preparedness Framework — its top cyber risk tier. In OpenAI's own testing, it found 81 previously unknown high-severity vulnerabilities in Firefox, most of which Mozilla has already patched, and built working local-privilege-escalation exploits for 948 Linux kernel flaws it reproduced (from 2,308 flagged and 1,724 reproduced), sharing 304 critical findings with Linux maintainers along with proposed patches. On its ExploitGym benchmark it scored 48.5%, up from 32.1% for the previous model. Access is limited to vetted customers in its Daybreak Red program.

Mary's take: read the numbers the careful way — these are OpenAI's own figures, not independently replicated yet, and 948 working exploits is not the same as 948 brand-new bugs. But the direction is unmistakable, and it is the same one from two weeks ago: AI is finding flaws faster than humans ever could. The finding is not the risk. The patching speed is. Ask your team one question this week: when AI finds a flaw in something we run, how fast does the fix actually ship?

3. The trusted tool was the trap — check your Cursor extension

This one started October 29 and landed in our week on November 2, when Cursor confirmed both tampered versions. For about 11 hours on October 29, maliciously modified builds of Anthropic's Claude Code extension — versions 2.1.312 and 2.1.313, for macOS and Windows — were served from Cursor's own extension marketplace. The tampered code tried to plant a persistent SSH backdoor: an attacker's key, a reverse tunnel, persistence on the machine. Cursor says at most 5,000 users installed the bad version, and it will contact them directly. The clean version is 2.1.314 — and uninstalling the extension alone does not remove the backdoor, so Cursor published cleanup scripts for each platform.

Mary's take: this is the supply-chain lesson in its purest form. Nobody installed something shady. They updated a tool they trusted, from a store they trusted, and the store was the problem. If your developers use Cursor, the check takes a minute: did anyone install or update Claude Code on October 29, and are they on 2.1.314 or later, with the cleanup run if needed? Trust, but verify the update.

4. The receipts got real: watermarks, proofs, and a rewritten rulebook

This was the week AI started having to show its work in a new way. On November 3, Anthropic extended its invisible text watermark to older Claude models — a statistical pattern in word choice, no extra characters, no opt-out — and xAI followed on November 7 with the same kind of mark on seven Grok API models. The trigger is the EU AI Act's transparency rule, and Anthropic applies the mark worldwide. Anthropic also published FlashProve, a research prototype that uses cryptography to prove which model actually produced a given output. And OpenAI rewrote its Model Spec — the public document that says how its models should behave — around five core values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care.

Mary's take: last week I said the next phase runs on evidence, not claims. This week the labs started building the evidence in. A watermark is not a guarantee — Anthropic is frank that it can only show Claude was likely involved, and a heavy rewrite can strip it — but it is a start. The question for the rest of us is the same one, just sharper: when an AI hands you an answer, a document, or a decision, can anyone prove where it came from?

5. The week ended at the top — and in the jobs numbers

On Sunday, November 8, Nvidia's board opened an independent investigation into reports that CEO Jensen Huang was targeted by a foreign intelligence operation, after reports of his relationship with a businesswoman U.S. intelligence agencies concluded was a spy for Beijing. Huang admitted he was unfaithful to his wife and said he supports the inquiry; the board said it has no reason to believe national security or Nvidia secrets were compromised, and Huang continues as CEO while independent directors and outside counsel run the review. House committee chairs have asked him for a closed-door briefing on November 18. The espionage allegations themselves remain reports, not established fact — and that distinction matters.

And in the same week's jobs data, professional and business services in the U.S. added 70,000 jobs in October — the largest monthly increase since July 2022, per the Wall Street Journal. One month is not a verdict on AI and jobs, and I am not going to pretend it is. But it is worth writing down, because the headline version of this story is always "AI is taking the jobs," and the numbers, this month at least, told a different story.

Mary's take: governance is not a side story anymore — it is the story, all the way up to the most important company in AI. And on jobs, watch the hiring, not the headlines. The question I keep coming back to is not whether AI changes work. It does. It is who is still in the room, with judgment, when it does.

The one thing to remember

This was the week the receipts arrived from every direction at once: an examiner's framework, a model's exploit count, a marketplace cleanup, a watermark, and a board investigation. The hype phase runs on claims. This phase runs on proof — who can show it, who can check it, and who answers when it is missing.

This week's one assignment

Pick one AI tool your team already uses and ask three questions about it: can anyone prove what it did, how fast does a flaw it finds get fixed, and who answers when it acts on its own? If you cannot get a clean answer to all three, you just found your next conversation — not your next purchase.

See you next Tuesday. Same time, fewer tabs.