| ← All issues | |
| The Diff No. 037 Monday, 9 November 2026 · AI dev news, only what is verified |
 |
Goldman Sachs pulled Claude Code and GitHub tools from its bankers and traders - Bloomberg reported the October guidance, and the same bank is a lead underwriter on Anthropic’s IPO. Plus: Nvidia’s board investigates Jensen Huang, in the board’s own words. Bloomberg, Nov 2, via people familiar: Goldman restricted AI coding tools to its ~12,000 engineers, cutting off traders and bankers who had been building with them - weeks before this week’s saved-secrets scare, and weeks before Goldman takes Anthropic public. Also today: Nvidia’s board opens an independent investigation into CEO Jensen Huang after reporting that he was targeted by a foreign intelligence effort - what the board confirmed, what outlets allege, and what Huang’s draft email says; and a developer plants 12 fake secrets to test the scanners behind the 1,122-file story. The bank’s quiet retreat, the board’s Sunday statement, and the scanner test you can rerun. |
IN THIS ISSUE  | 01 🏦 Goldman pulled Claude Code from its bankers and traders - weeks before taking Anthropic public Bloomberg reported Nov 2 that October internal guidance restricts Anthropic’s Claude Code and Microsoft’s GitHub coding tools to Goldman’s engineers, after a productivity and risk review. Strats keep access. Goldman is a lead underwriter on Anthropic’s reported IPO. | | 02 🔍 Nvidia’s board is investigating Jensen Huang - what the board said, what outlets allege Nvidia’s official statement, Nov 8: the board opened an independent investigation after reporting that CEO Jensen Huang "has been the target of a foreign intelligence effort." Axios obtained a draft email in which Huang apologizes to employees for an affair. Whether the woman was actually intelligence is unproven. | | 03
🧪 "I planted 12 fake secrets. 4 beat both scanners" - the 1,122-file story, reproduced Developer frank chu reran the checks behind Apiiro’s Claude Code findings: Gitleaks 8.30.0 caught 6 of his 12 planted secrets, TruffleHog 3.91.0 caught 7, and 4 - every one a bare password with no provider format - beat both. His test, his numbers. |
|
🔥 VIRAL The loud ones on Hacker News. Numbers are a morning snapshot. |  01 Bloomberg via thestateofai.com, Nov 2 · HN: 669 points at 06:26 IST 🏦 Goldman pulled Claude Code from its bankers and traders - weeks before taking Anthropic publicBloomberg reported Nov 2 that October internal guidance restricts Anthropic’s Claude Code and Microsoft’s GitHub coding tools to Goldman’s engineers, after a productivity and risk review. Strats keep access. Goldman is a lead underwriter on Anthropic’s reported IPO. SCENE In July, Goldman’s CIO Marco Argenti was publicly praising "vibe coding" - traders and bankers building their own tools with AI and handing them to engineers to harden. Sometime in October, the guidance changed. By the time Bloomberg reported it on Nov 2, the people closest to the trading floor had lost the tools entirely. THE THING
The reported facts, per Bloomberg’s unnamed sources: Claude Code and GitHub coding tools are now restricted to the bank’s roughly 12,000 engineers (of ~47,000 staff); strats - the hybrids across trading, quant research and engineering - keep access. Traders had used the tools to monitor client flows, model market moves and assemble materials. Hong Kong bankers had already been barred from Claude in April. Goldman spokesperson Nick Carcaterra’s only public line: the bank makes AI tools available "to determine the best use cases" across the firm. The kicker the bank did not put in that statement: Goldman is a lead underwriter on Anthropic’s reported pre-Thanksgiving IPO. All of this predates this week’s Apiiro saved-secrets story (No. 036) - Goldman’s risk review concluded before the public scare, which is exactly what makes it a signal rather than a reaction. Bloomberg’s reporting, their sources. ✂️ STEAL IT Steal the governance timeline: a top-tier bank looked at non-engineers running code-writing agents on machines with client data, and decided the permission model was "engineers only." If your org lets analysts, sales or ops run coding agents today, expect the same review - and steal Goldman’s middle tier: a strats-like group that keeps access because it sits between business and engineering. Watch VERDICT Read the Bloomberg-derived reporting and watch whether other banks follow - one bank’s October guidance is a policy, three banks’ is an industry standard. It fails you if you want Goldman’s own account; only the spokesperson line exists.
State of AI writeup of the Bloomberg report · Bloomberg Law summary · Quartz coverage · HN discussion |
02 nvidianews.nvidia.com statement + Axios, Nov 8 · HN: 449 + 417 points at 06:26 IST 🔍 Nvidia’s board is investigating Jensen Huang - what the board said, what outlets allegeNvidia’s official statement, Nov 8: the board opened an independent investigation after reporting that CEO Jensen Huang "has been the target of a foreign intelligence effort." Axios obtained a draft email in which Huang apologizes to employees for an affair. Whether the woman was actually intelligence is unproven. SCENE Sunday evening, the most valuable company in the AI supply chain posted a five-sentence statement. Not a leak, not a rumor - the board itself, announcing that it is investigating its own founder-CEO, and adding that it "has confidence in Jensen." THE THING
Separate the three layers, because they are not equally solid. Board-confirmed: an independent investigation is underway, run by a committee of independent directors with outside counsel; the trigger was "recent reporting" that Huang was the target of a foreign intelligence effort; the board "has no reason to believe that either national security or NVIDIA secrets were compromised"; Huang "fully supports the review." Reporting allegation: outlets including Axios and CNBC describe the relationship as with a "suspected" or "alleged" Chinese spy - no outlet has published evidence she was intelligence. Private draft attribution: Axios obtained what it describes as Huang’s draft email to employees - "I owe my family and all of you an apology... I was unfaithful to my wife" - and Huang has not publicly confirmed sending it. The developer consequence is real and separate from the scandal: Huang has personally pushed to loosen U.S. export rules on advanced AI chips to China, CNBC reports Congress is already circling, and any leadership instability at Nvidia lands directly on GPU supply, pricing and the export policy every AI team builds against. ✂️ STEAL IT Steal the reading discipline: when a story has an official statement, an allegation, and a private document in the same headline, price each layer separately before you repeat any of it. The board’s five sentences are worth more than fifty derivative articles. Watch VERDICT
Read the board statement itself - five sentences, linked below - and the Axios piece for the draft-email wording. It fails you if you want resolution; the investigation has no timeline and the board promises only "further updates as appropriate." Nvidia board statement · Axios: the draft email · CNBC: Congress scrutiny · AP report · HN discussion |
03 dev.to, Nov 8 🧪 "I planted 12 fake secrets. 4 beat both scanners" - the 1,122-file story, reproducedDeveloper frank chu reran the checks behind Apiiro’s Claude Code findings: Gitleaks 8.30.0 caught 6 of his 12 planted secrets, TruffleHog 3.91.0 caught 7, and 4 - every one a bare password with no provider format - beat both. His test, his numbers. SCENE After reading Apiiro’s claim that 1,122 public settings.local.json files hold secrets (No. 036), frank chu did the thing readers always ask for: he planted 12 fake secrets in Claude Code settings files and ran the exact scanner versions Apiiro used. A placeholder password and a secrets-free file, his controls, produced zero false positives. THE THING
The result that should worry you: the four secrets that beat both scanners were all bare passwords - no sk- prefix, no AKIA, no recognizable provider format. Pattern-based scanners look for the shapes of known credentials; a password like the one you typed into a psql command has no shape. Chu also cross-checked Apiiro’s arithmetic: Gitleaks flagged 74.4% of the real secrets Apiiro found, TruffleHog 71.6%, either tool 84.4% - and he caught one inconsistency, a per-type chart summing to 1,384 versus the headline 1,379. His independent floor via Sourcegraph’s public index: 928 copies of the file across 841 repositories, skewed to starred repos, so a floor not a recount. And in a second post he rebuilt Anthropic’s memory filing mechanism in 80 lines of Python with the model stripped out entirely - the 413-files-vs-9 gap from No. 036 survives, "mechanical enough to measure with grep." His tests, his planted secrets, his rebuild. ✂️ STEAL IT Steal the test itself: it takes an afternoon. Plant a dozen fake secrets in a scratch repo - some with provider formats, some bare passwords - and run whatever secret scanner your CI uses. What slips through is your real coverage, and bare passwords are the blind spot to close with process, not patterns. Try today VERDICT Rerun his protocol against your own CI scanner; the gap he found is in the tool category, not one vendor. It fails you if your threat model is only provider keys - then the scanners’ 74-84% is your number instead.
The scanner test · The memory-mechanism rebuild · No. 036: the original Apiiro story |
📊 THE ATTENTION GAP HN points at 06:26 IST. Attention, not quality. | Goldman pulls Claude Code from bankers and traders | | 669 HN | | Claude Code saved-secret files (Apiiro, No. 036) | | 596 HN | | Nvidia board investigates Jensen Huang | | 449 HN | | Huang’s letter to staff (Axios) | | 417 HN |
|
ALSO ON THE RADAR Langford names the philosophy that killed AI Frontiers: "scaling, and aggressively not research." A primary-source detail worth adding to No. 036: in his own post, Langford writes that AI Frontiers "were reorged into Microsoft AI" in early September and that the philosophy of the ex-Inflection leadership of Microsoft AI "is incompatible... scaling, and aggressively not research." His account, his words; Microsoft still has not posted its own. Langford’s post → Declarative Dragging: say what a drag should become, not how it should move. Joshua Horowitz, Andrew Blinn and Jeffrey Heer’s system speculatively renders the future states you declare for each draggable element, then works out geometry, feedback and animation itself. 388 HN points. Their project, their demo. The project → Anthropic’s dreaming numbers get their first outside poke - at the mechanism, not the eval. The frank chu rebuild in R03 strips the model out and reproduces the 413-vs-9 filing failure; the headline 13%-to-48% dreaming eval remains Anthropic’s own, unreplicated. The rebuild → |
🗂️ THE BORING STUFF Nov 8 · Outlook.com is quietly rejecting email from domains without DMARC. mailroute.net’s blog, 277 HN points. If you run any newsletter or transactional mail, check your DMARC record today; Buttondown handles ours. The post → Nov 8 · Herb Sutter: "When a language’s creator and its owners strongly disagree" (PDF). ISO C++ paper P4454R0, 84 HN points. Governance-of-languages reading; title level unless C++ standardization is your job. The paper → Official pages and named outlets read this morning; the BORING items were read at title level. Nothing installed or run; their numbers are theirs. |
🧪 ONE THING TO TRY TONIGHT Plant your own secrets, then count what slips. Thirty minutes. First, the scanner test: create a scratch repo, write a dozen fake secrets into a Claude Code settings file - a few with real provider formats (sk-..., AKIA...), a few bare passwords like the ones you would actually type into a psql or sshpass command - and run your CI’s secret scanner on it. Bare passwords are the blind spot frank chu measured; find out if they are yours. Second, the Goldman test: list everyone in your org who runs a coding agent today and is not an engineer, then ask what data those machines can reach. If the answer is "client data and we have not reviewed it," you are pre-October Goldman. Fix the permission model before someone’s risk review - or someone’s scanner - does it for you. |
What this issue could not see. HN and GitHub were swept overnight and this morning, and primary sources were read: the Nvidia board statement in full at nvidianews.nvidia.com, Axios’s report in full, thestateofai.com’s writeup of the Bloomberg report in full, both frank chu posts in full, hunch.net in full, declarative-dragging.github.io, and the HN front page at 06:26 IST. The Goldman facts are Bloomberg’s reporting via unnamed sources, published Nov 2; Goldman has said only that it makes AI tools available "to determine the best use cases." On Nvidia: the investigation, the independent committee, the no-reason-to-believe-compromise line and the confidence in Huang are board-confirmed; the "suspected/alleged Chinese spy" framing is outlet reporting without published evidence; the apology quotes come from a draft email Axios says it obtained, which Huang has not publicly confirmed sending. Bloomberg Law’s headline ties the probe to the admitted affair; the board statement does not. frank chu’s scanner and memory numbers are his own tests. One outlet (dailyaibrief.com) claims "2,539 developer secrets" - that contradicts Apiiro’s own 1,379 figure and we do not repeat it. Reddit remains absent by policy; X blocks anonymous access and was not consulted. Nothing was installed or run. VIDEO OF THE DAY is omitted for the fourth day running: nothing fresh, relevant and verifiable on today’s stories turned up on YouTube - we omit rather than pad. Deduped against every published issue, No. 001-036, via the kit shipped-index and the live home archive: the Goldman restriction and the Nvidia investigation have never been covered; the Langford item is new primary-source detail on No. 036’s story, not a re-report; the scanner test is external follow-up to No. 036’s lead. |
| Images: credited in each report; source screenshots and announcement artwork, not our tests. |
|