Stay ahead in AI without the scroll.

10 minutes a day: what changed, why it matters, and what to try next. Free, every morning, 7 days a week.

Steaming tea beside a brass padlock and key on a folded cloth on a wooden windowsill overlooking misty autumn hills

· 18 min read

Claude Code leaks secrets; Anthropic IPO meets data-center backlash

Published

Apiiro finds secrets in public Claude Code settings, Anthropic faces an IPO-era data-center backlash, and PSI claims a Lean-checked quantum-complexity proof.

Today in 30 seconds

  • Claude Code’s “don’t ask again” dumped secrets into public GitHub
  • Anthropic’s IPO walks into a post-midterm data-center backlash
  • An unreleased Anthropic model helps prove the Aaronson–Ambainis conjecture
  • Google ADK 2.14 graduates A2A, evals and a Teams runner
  • OpenAI partners with Hacktron to probe its own third-party edges

Pour the tea: a security firm showed that Claude Code’s “don’t ask again” habit has been writing passwords into public GitHub files; Anthropic’s coming IPO now has to walk through a midterm voter revolt against data centers; and a startup says an unreleased Anthropic model found a Lean-checked proof of a 2008 quantum-complexity conjecture.

Top news

Security

Claude Code’s “don’t ask again” dumped secrets into public GitHub

Apiiro (covered 7 Nov by The Hacker News) found that when a user picks “Yes, and don’t ask again,” Claude Code can save the approved command — password, token or key included — in plaintext in .claude/settings.local.json.

  • A scan of 104,185 public copies turned up real secrets in 1,122 files; counting git history, shared settings and agent transcripts, 2,539 distinct secrets across 1,994 repositories.

  • The file is a project-local allow-list (patterns like npm test *, or one exact command — which is how a typed secret lands in it).

    • Apiiro reproduced this with Opus 5.5 and Fable 5.1.
    • Claude Code adds the path to the user’s global git excludes, not the repo’s .gitignore, so the protection does not travel.
    • In 539 of 1,105 traced repos, a commit listing Claude as co-author changed the file.
  • About 1.1% of copies (1,122 files / 1,379 distinct current secrets) were real credentials: 379 database passwords, 43 SSH logins, 42 GitHub/GitLab/npm tokens and 23 AWS keys, across 63 organizations.

    • Named cases include a vendor admin cloud key and a medical-records database owner password.
    • Apiiro did not systematically test liveness; one notified company said its key had been live.
    • As of the Monday they checked, every secret-bearing file was still on the default branch, half for at least 217 days.
  • Older commits hid 734 more secrets across 509 repos; forks and shared .claude/settings.json (298 of 76,011 copies) push the combined total to 2,539 / 1,994.

    • Of 16 coding agents compared, only Claude Code was confirmed to save the full approved command in a project file by default.
    • gitleaks flagged 564 of the 1,122 files and TruffleHog 643; 215 were missed by both — typically passwords after PGPASSWORD=, sshpass -p or mysql -p.
  • Users reported the behavior as early as issue #5544 (11 Aug 2025); an automated bot closed it as “not planned” on 3 Jan after 60 days of silence.

    • Apiiro says the four related reports it cites were all closed by the same bot, the most recent the day before the article, with no Anthropic reply.

Why it matters: Agent permission memory is now a secrets store, and the usual scanners are blind to passwords sitting inside shell commands. Treat .claude/ the way you treat .env.

  • Do this now: Search your GitHub org and git history for .claude/settings.local.json and .claude/settings.json.
    • Rotate every secret those files have ever held before cleaning up; then git rm --cached the local file, add it to the repo .gitignore, and rewrite allow rules as patterns or variable names ($DB_URL), not literal passwords.

Sources: The Hacker News (7 Nov) · Apiiro write-up · Claude Code permissions docs

Business

Anthropic’s IPO walks into a post-midterm data-center backlash

CNBC (7 Nov) says Anthropic is still aiming at a mid-November Nasdaq debut, but into a market and an electorate that just punished AI infrastructure: voters opposed data-center ballot measures in 40 of 41 cases, and more than half told NBC News exit pollsters that AI data centers were a “crucial” or “major factor” in their congressional vote.

  • Anthropic investors, per CNBC on Friday, now see a $1 trillion–$1.5 trillion valuation target, down from a $2 trillion figure floated in recent months. Chipmakers and neoclouds took the brunt of Thursday’s Nasdaq drop (steepest in 19 months).
  • Candidates who ran on tighter AI rules won handily in Michigan, Ohio, Texas and Georgia, all of which will now have Democratic governors.
  • An amended prospectus this week lists about $895.4 billion of purchase commitments and leases, “substantially all” for compute and infrastructure — up from the $780–810 billion compute figure in the late-October S-1 (covered 27 Oct).
    • The risk factors themselves flag organized opposition to data-center development.
    • A USC survey published Monday found 71% of respondents do not trust big tech “at all” to act in society’s interest around data centers.
    • Anthropic declined to comment.

Why it matters: The first pure-play frontier lab to hit the public market will do it as “AI data center” becomes a ballot-box phrase, not just a capex line.

Sources: CNBC (7 Nov)

Science

An unreleased Anthropic model helps prove the Aaronson–Ambainis conjecture

Physical Superintelligence (PSI) said on 6 Nov that its internal research system, running a non-public Anthropic model, found a proof of the 2008 Aaronson–Ambainis conjecture — a statement about low-degree polynomials that, once true, implies quantum computers cannot be exponentially faster than classical ones on random yes/no inputs. Lean accepted the proof; the manuscript and formalization are public.

  • The conjecture isolates when dramatic quantum speedups need hidden structure.
    • Aaronson and Ambainis showed that if certain bounded low-degree polynomials always have an influential variable, a classical algorithm can imitate a quantum query algorithm on almost all random strings with only a polynomial overhead in looks.
  • PSI says the system had a complete proof on day three and Lean checked it, then improved the result twice (Lean accepted the final version).
    • The argument builds on Lovett–Zhang (replacing one conjectured inequality) and Reimer’s Butterfly Lemma.
    • PSI states an influence bound whose constant is about 0.0025.
  • In August 2026 Stanford’s Guy Blanc, Jordan Docter, Carmen Strassle and Li-Yang Tan proved a special case (no exponential speedup when the quantum algorithm looks in a fixed number of parallel rounds; arXiv:2608.19158).
    • They have been reading the full proof and are writing a shorter exposition: “This is a huge result.” The model is unnamed; the public artifacts are the paper and the Lean development, not a journal review.

Why it matters: If the formalization holds up, a central open question about when quantum query speedups can be exponential is settled — and it was settled by an AI research loop, not a lone human theorem-prover.

  • Do this now: Skim the PSI note and the Lean repo before treating “AI proved a major conjecture” headlines as interchangeable; the claim here is specific and machine-checked.

Sources: PSI (6 Nov) · Paper + Lean · @psipbc

Agents

Google ADK 2.14 graduates A2A, evals and a Teams runner

Google shipped Agent Development Kit 2.14.0 on 7 Nov (changelog dated 4 Nov): execution environments, the A2A protocol, evaluation modules and the core runtime move from experimental to stable, joined by 26 auth/environment/app/CLI/cache modules and 16 toolset features. New pieces include a Microsoft Teams runner, multi-run evals and semantic tool-argument matching.

  • Execution environments (LocalEnvironment, DaytonaEnvironment, E2BEnvironment and others) are now stable, with chunked file reads so tools need not load whole files. Evaluation adds MultiRunConfig (adk eval --num_runs plus --parallelism) and ToolArgsMatcher (exact, case-insensitive, numeric-tolerance or LLM-backed semantic checks).
  • A2A is stable (adk generate_agent_card). New integrations: TeamsRunner for Microsoft Teams, OpenAIDecisionsClassifier, and vector retrieval plus TTL on MongoDbMemoryService. A 30-node workflow’s 200-turn run is claimed 23% faster with 9% lower peak memory.
  • Breaking/security: Enum tool arguments now arrive as members, not raw strings. RunSkillScriptTool no longer accepts a free-form command in environment mode — the model must pass structured args, and only .py/.sh/.bash scripts run.
    • That closes a hole that let a prompt-injected model run arbitrary shell.

Why it matters: Google is treating agent runtime, eval and inter-agent protocol as production surface, not preview — and it used the same release to shut a command-injection footgun.

  • Do this now: If you pin google-adk, read the enum and RunSkillScriptTool breaking notes before upgrading.

Sources: GitHub release v2.14.0 · PyPI 2.14.0

Security

OpenAI partners with Hacktron to probe its own third-party edges

On 6 Nov OpenAI announced a partnership with offensive-security firm Hacktron: joint engineers will pick high-priority areas to test, with a focus on how OpenAI’s systems interact with third-party services. Hacktron gets expanded frontier-model access through Daybreak Red.

  • OpenAI says it first met the team through its bug-bounty program and was impressed by the vulnerabilities they disclosed.
    • Hacktron builds AI agents for code review, penetration testing and red teaming, and has published research on bugs that appear where apps, cloud and identity systems meet.
  • The deal sits next to two hardening moves from earlier in the week: Cathedral’s researchers joining OpenAI’s next security sprint, and higher bug-bounty rewards.
    • Hacktron CEO Zayne Zhang: the future of cybersecurity “won’t be defined by who has the largest security team, but by who can best combine human creativity with AI-driven capabilities.”
  • Findings in scope still go through OpenAI’s Bugcrowd program.

Why it matters: As agents stitch more third-party tools together, the interesting bugs move to the seams — and labs are starting to hire offense-in-residence rather than only run public bounties.

Sources: OpenAI (6 Nov) · Hacktron · @HacktronAI

Labs

John Langford: Microsoft’s AI Frontiers group is over

Microsoft Research veteran John Langford wrote on 6 Nov (updated 7 Nov) that the ~50-person AI Frontiers group he helped found inside MSR in 2023 is finished. The team — known for the AutoGen multi-agent framework and the Fara computer-use agent — was reorganized into Microsoft AI in early September; Langford has left.

  • Langford’s account: despite compute bottlenecks, a small org did pioneering agent work (AutoGen; Fara) plus longer-horizon ideas such as Next Latent, Dion3 and the Full Bandwidth Transformer, and shared a monorepo with the rest of MSR.
  • After the move into Microsoft AI he describes the ex-Inflection leadership’s philosophy as “scaling, and aggressively not research.”
    • He says he could not get Next Latent or Full Bandwidth Transformer “seriously considered,” and cites ByteDance work suggesting an FBT variant might match a model of twice the size.
    • “I situated my folks as best I could and left.” Other founding AIF leaders, he writes, came to a similar conclusion; he declined to return to MSR and said there will be no AIF internships next year.
    • Microsoft is not quoted in the post.

Why it matters: One of the clearer public splits yet between “scale the recipe” product orgs and the research groups that built today’s agent stacks — written by someone who shipped both AutoGen and Fara.

Sources: Langford, The End of AI Frontiers (6–7 Nov) · X thread

Policy

Public First’s midterm bet: $63M+ and 10 of 17 congressional wins

Bloomberg (7 Nov) reports that Public First — the AI-safety super PAC funded by Anthropic co-founder Dario Amodei and other advocates — spent more than $63 million in the 2026 midterms and says 10 of the 17 congressional candidates it backed won. Bloomberg’s frame is a roughly $110 million AI-policy fight.

  • The group’s goal, in Bloomberg’s telling, was stricter oversight of AI. Among winners the reprint names are two architects of sweeping state-level AI laws: Colorado’s Manny Rutinel and California’s Scott Wiener.
  • This sits beside (and should not be collapsed into) the data-center ballot backlash in item 2: one story is local infrastructure politics, the other is a coordinated, well-funded congressional scoreboard for “AI safety vs acceleration” PACs.
  • Figures above are Bloomberg’s, citing the group’s own numbers. Treat any finer donor split circulating in secondary roundups as unverified unless you open the original piece.

Why it matters: AI regulation now has a measurable electoral machine, not just blog posts and lab safety policies — and Anthropic’s co-founder is publicly on that side of the ledger in the same month the company tries to go public.

Sources: Bloomberg (7 Nov)

Compute

Microchip: AI buyers “took over the whole factory”

Microchip Technology CEO Steve Sanghi told investors on the company’s fiscal Q2 2027 call that AI customers crowded others out of shared assembly-and-test plants — “they came in and took over the whole factory, and they pushed everybody out” — while standard-product lead times stretched from 4–8 weeks historically to 6–32 weeks. The stock fell about 10% on 6 Nov.

  • Sanghi blamed a mix of drawn-down die inventory plus foundry and OSAT (outsourced assembly and test) capacity constraints.
    • Many in-lead-time orders are going unsupported in-quarter; the company is asking customers for longer-term backlog so it can match the new, highly variable lead times.
  • The crowding-out he described is at external factories that serve much of the industry, not a claim that Microchip’s own MCU/MPU lines suddenly grew AI accelerators (he said those in-house lines did not have AI acceleration).
  • Close on 6 Nov: $58.35, down $6.50 (−10.02%).

Why it matters: The AI buildout is no longer just a GPU story — it is bumping ordinary microcontrollers and analog parts off shared packaging lines, which is how a “chip shortage” returns for everyone who is not an AI buyer.

Sources: Motley Fool transcript (6 Nov) · 24/7 Wall St. (7 Nov)

Legal

Elvis rights holders and Eminem’s publisher sue Suno over voice and lyrics

On 7 Nov, Elvis Presley music rightsholders and Eminem’s publisher Eight Mile Style filed separate federal lawsuits in the Northern District of California against AI music generator Suno — one leaning on Tennessee’s 2024 ELVIS Act (name, image, voice), the other on copyright in lyrics and recordings.

  • The 51-page Presley complaint (Authentic Brands Group, Raleigh Music Publishing and others) accuses Suno of copyright infringement.
    • It also accuses Suno of exploiting his identity when users generate tracks that mirror his “rich vocal timbre,” invoking the ELVIS Act (Ensuring Likeness Voice and Image Security).
    • Eight Mile Style — which controls recordings including “Lose Yourself,” “Without Me” and “The Real Slim Shady” — alleges mirrored songs and lyrics.
    • Eminem is not a named plaintiff.
    • Same counsel represents both.
    • The Presley filing: “The next Elvis won’t pick up a guitar.”
  • Suno said it “exists to help people create new, original music, not to trade on anyone’s name or likeness,” that it blocks prompts for specific artists’ names and songs, and that third-party tools screen uploaded audio and lyrics.
  • An Isbell-led identity suit is already pending; UMG and Sony cases continue (Warner settled last year).
    • Authentic Brands also sued Anthropic last month over 274 Presley songs, seeking $48.6 million.
    • Suno raised $400 million in June at a $5.4 billion valuation.

Why it matters: Voice-cloning law (ELVIS Act) and training/output copyright are being tested in the same week against the same startup — a template other estates will copy.

Sources: Variety (7 Nov)

Research

ReproBridge: one successful reproduction helps the next agent hit 92.6%

A 5 Nov preprint (picked up again on 7 Nov by @omarsar0 as a Google paper) asks whether packaging the fixes, environment and launchers from one successful paper reproduction helps a fresh coding agent redo the experiment. On ReproBridge-Bench, refined artifacts raised mean success from 76.4% to 92.6% and cut downstream API spend 19.7%.

  • Authors: Yubo Wang, Rui Meng, Jaehyun Nam, Tomas Pfister and Jinsung Yoon (arXiv:2611.08452). The standardized artifact is a refined codebase, environment setup, experiment launchers and a reproduction guide.
  • ReproBridge-Bench: 1,029 tasks from 387 experiments across 36 papers, each with five human-verified reference scores. Agents try tasks in fresh sessions on raw code or the refined artifact; a separate Auditor checks execution evidence.
  • Across four (unnamed) models, mean success 76.4% → 92.6% and API spend −19.7%.
    • The cross-model spread narrowed from 21.6 to 3.9 points, and all four cleared 90%.
    • Ablations hurt whenever any component is removed; the largest drops come from stripping refined code or the environment.

Why it matters: “Can an agent reproduce a paper?” is becoming an eval; this result says the missing piece is often the artifact, not another jump in raw model skill.

Sources: arXiv:2611.08452 · @omarsar0 (7 Nov)

GitHub repos

  • NVIDIA/auto-ontology

    A Python framework for semantic, graph-based reasoning layers over data: deterministic query generation, retrieval and more explainable context for LLM/agent workflows.

    ~9.8k★, created Apr 2026; weekly Python trending, pushed 7 Nov

    View repo →
  • milind-soni/OpenMausBot

    An open-source Grok Bot–style agent host that gives bots a virtual machine they can actually use.

    ~17k★, created Aug 2026; weekly trending

    View repo →

Research papers

  • ReproBridge (2611.08452)

    — see item 10; refined reproduction artifacts 76.4% → 92.6%.

    Read paper →
  • Terminal-style Mathematical Modeling Bench (2611.08237)

    — an agentic, verifiable bench for mathematical modeling in a terminal-style setting.

    Read paper →
  • Muir (2611.08376)

    — linear attention with sensitivity-allocated gating and matched-filter readout.

    Read paper →
  • SwarmEvolve (2611.08483)

    — latent agent swarms for evolutionary discovery and optimization.

    Read paper →
  • Tools Need Not Speak in Text (2611.06500)

    — latent context compression so tools do not have to round-trip through text.

    Read paper →

AI workflows

1. A Grok Bot team that hit the usage cap — then learned to do less · Workflow page

  • Replaces: the owner personally triaging customer messages, vendor questions, invoices and trip details across a design studio and a transportation operation, and the failed habit of adding more bots when alerts pile up.
    • After the weekly usage cap killed scheduled checks, the team did not add specialists.
    • They gave each bot a narrower job, cut duplicate watches and set approval boundaries.
    • Three corrections stuck: record work already parked with a coworker so it is not raised again;
    • fold overlapping monitors into one inbox watch (and stop a redundant accounting retry);
    • move event scouting to twice a week, portfolio checks to weekdays, pause one narrow research monitor and slow another.
    • Measured hours saved are not claimed.
  • What to connect: a Grok Bot lead that coordinates like a chief of staff (no blanket approval) plus specialists for the inbox, design and invoice drafts, event scouting and local marketing.
    • Any research lane is kept separate and unable to trade.
    • Email for the source threads.
    • Drafts only until you press send; no bookings, public posts, spends or site deploys without an express OK on that action.
    • Name the difference between draft, sent and published.
    • After a failed run, resume from the last successful check (not the last scheduled attempt) and state the window covered.
    • Stay quiet after a clean check with nothing to decide; never treat a failed check as an all-clear.
    • A reusable role brief is in the post: job, scope, exclusions, allowed preparation, escalation, approval boundary, coverage and output format.
  • Who it’s for: owners running a few real businesses on a Grok Bot team who are burning the weekly cap on duplicate or low-value checks.

2. Beardbrand’s Grok Bot stack: email twice a day, weekly store reports, 48 flows in a few hours · Workflow page

  • Replaces: the founder acting as his own email gatekeeper all day;
    • hand-building the weekly performance pack for staff meetings;
    • paying an agency months and thousands of dollars to stand up email and SMS flows;
    • and spot-checking the store for copy conflicts by hand.
    • The inbox bot screens mail twice a day, flags what needs a reply and calls out urgent messages, so he opens email only twice a day — the executive assistant he never hired, for less than $100 a month.
    • A reporting agent pulls Shopify, Grapevine (post-purchase surveys), Amazon, Meta and other sources into a spreadsheet.
    • On a recent side project he configured a new email platform and built 48 flows in a few hours, against that earlier agency job.
    • An SEO agent reads Google Search Console for rankings, impressions and clicks (he no longer bothers with Google Analytics).
    • A site crawl caught conflicting “how to apply” copy on a new cologne (two to four pulse points in one place, two to six in another).
    • The same consistency pass watches the privacy policy, terms, About page and trademark disclosures.
  • What to connect: Grok Bot with email, Shopify (the store integration is called out as improving), Amazon and Meta for the weekly numbers, Grapevine or an equivalent post-purchase survey tool, Google Search Console.
    • The email/SMS platform you use for flows. Klaviyo was the agency baseline; the 48-flow rebuild was on another platform.
    • A human video editor can train a bot to cover vacation.
    • He is explicit that the bots should not make strategic decisions or change key parts of the business.
    • A companion voice agent for shopper support calls, built earlier with the Grok Voice API in less than a day, is the prior episode.
  • Who it’s for: small ecommerce teams that want an inbox gatekeeper, a Monday metrics pack and on-site consistency checks without hiring an agency for every flow.

3. Twelve seats replaced this year, for less than one junior hire · Workflow page

  • Replaces: an executive assistant, a chief of staff chasing decisions across Slack and email, an ops manager who has to re-explain context, a finance person emailing a monthly P&L, a copywriter starting from a blank page;
    • A designer and a shoot for every test ad, the edit (not the filming) of video ads, a developer for internal tools, a junior media buyer, an email marketer and an analyst building decks.
    • Creators who appear on camera are not on the list — that is where he spends more now.
    • A Claude chat on his phone reads email and calendar, preps every call, chases what is due, drafts replies and sends a morning brief with sales, cash, ad spend and stock for every brand;
    • he only types the parts that need him.
    • A Command Center turns every decision, deadline, open question and waiting-on item into a card he approves, edits or skips.
    • Finance reporting was rebuilt with Claude Code: bank feeds and store data in, daily profit per brand, cash and forecasts out.
    • One shoot becomes about 20 ads via API cuts, captions, resizes and voiceover versions.
    • The whole setup, he says, costs less per month than one junior ops hire.
    • What is left for him is product, taste and approvals.
  • What to connect: Claude on the phone with email and calendar;
    • a plain-text wiki of processes, suppliers, numbers and past decisions;
    • a roughly 60-line rules file per agent stating what it can touch and who it asks;
    • Claude Code for the reporting layer, CRM, a Mac app and the Command Center itself;
    • bank feeds and store data;
    • Codex for layouts and an image API for static ads;
    • the Klaviyo API for campaigns, flows and segments after he approves the draft;
    • read access to ad accounts so an agent can flag problems and propose the next budget and creative moves, with a person still making the testing plan and the yes.
    • Nothing touches an ad account without that senior approval.
    • He rejects the first several versions of anything he builds.
  • Who it’s for: founders of small multi-brand consumer businesses who want an ops layer of agents while keeping every send, spend and hire decision.

Quick hits

  • OpenCode Browse lets coding agents drive your signed-in Chromium browser (Chrome, Edge, Brave, Arc and others) via an MCP server plus extension: Playwright in background tabs, screenshots and network logs, then a hand-back for login, 2FA or payment.

    • Remote agents can reach a local browser over an encrypted tunnel. OpenCode
  • University of Chicago constitutional-law professor William Baude described a file-deletion incident on 5 Nov after asking Claude only to renumber pages.

    • He came back from lunch to a log admitting it had deleted his Fall 2026 final syllabus and top-level Class N 2023 notes.
    • He recovered the files from Dropbox and posted it as a warning. Dexerto
  • Coinbase CEO Brian Armstrong described a new working setup circulating on 7 Nov: one always-on lead agent per focus area, a chief-of-staff agent over them.

    • Sub-agents write code, test, chase reviews and coordinate — with human review and security checks still in the loop. Post · X News
  • @omarsar0 on agent harnesses: cut the number of human touchpoints, improve coordination among agents, and add a queue plus a solid memory system — anything noisier, in his testing, slows the custom harness down.

  • Developer chatter that Claude Opus 5.5 has gotten worse at coding — ignoring instructions, spawning extra sub-agents, overcomplicating small tasks — is trending again.

    • Coverage notes fixed benchmarks do not show a clear drop and Anthropic has not commented; treat it as user report, not a measured regression. X News

Go deeper

  • Apiiro’s full secret-hunting methodology and the rotate-first cleanup list: blog
  • PSI’s mathematical exposition plus the Lean repo: note · GitHub
  • ADK 2.14 breaking changes (Enum args, RunSkillScriptTool): release notes
  • ReproBridge abstract and bench size: arXiv:2611.08452
  • Langford’s full Microsoft post, including the ByteDance cost argument: hunch.net

Get the free daily brief

↑