
Leverlark · Issue #5 · Every week, the AI and automation plays that give a lean team the output of a big one.
Executive Summary
The model price ladder got clearer this week: a cheaper frontier tier sold as a planner, a near-free worker tier, and small “decision models” handling yes/no and labeling work for a fraction of the cost. The lean-team play is to route each job to the cheapest model that can do it. But first, spend 30 minutes on plumbing: a three-day Outlook glitch showed what happens to domains with no DMARC record, agent settings files are leaking secrets into public repos, and three deadlines land before New Year.
Top 3 News Highlights
1. Outlook bounced mail from domains with no DMARC record. It was a bug, so fix it anyway
From Nov 6 to Nov 9, Outlook.com, Hotmail and Live rejected some mail from domains that publish no DMARC record, even when SPF and DKIM passed. The bounce read 550 5.7.1 Message rejected: Message lacks SPF/DKIM/DMARC Auth Records, and it hit low-volume senders, not just the 5,000-a-day bulk senders covered by Microsoft’s 2025 rule. Microsoft has since confirmed it was a fault: in incident EX1545193 it said a spam-rule update “caused email messages to be incorrectly marked and blocked,” rolled it back by Nov 9, and says senders don’t need to change anything. Still, deliverability tester GlockApps found that while the fault lasted, any DMARC record, even p=none, stopped the bounces. For three days, receipts, password resets and outreach from domains without one just vanished.
The play: Add one TXT record at _dmarc.yourdomain.com with the value v=DMARC1; p=none. It takes two minutes, it’s cheap insurance against the next glitch, and you can tighten the policy later once reports come in. The Register · GlockApps tests · MailRoute · Hacker News discussion
2. Route the work: the three-tier model stack
Both Anthropic and Google now sell their top model as a planner that directs cheaper workers. Anthropic released Claude Fable 5.2 (and Mythos 5.2, the same model for vetted orgs) on Nov 9, with up to 30% faster output than 5.1. Anthropic still recommends Opus 5.5 as the daily driver and pitches Fable as an “Advisor” or “Orchestrator”. Google opened Gemini 4 Carbon to paying API developers on Nov 10 and previewed /swarm in Antigravity, where Carbon plans and audits while lighter models do parallel work. Underneath, Meta’s Muse Ember (Nov 9), its smallest model, targets tool calls, coding subtasks, routing and extraction, with a 1M-token context.
Frontier planner: Claude Fable 5.2, $8 / $40 per M tokens in/out (fast mode $16 / $80). Independent check: Artificial Analysis ranks it #1 of 74 (index 65), but at $26.54 and 44.6 minutes per index task.
Frontier planner: Gemini 4 Carbon, $2 / $10 intro, $4 / $20 from Dec 1. Budget at the $4 / $20 rate. Google's "#1 on CursorBench 4.0" and ">2X lower cost per task" claims are its own.
Cheap worker: Meta Muse Ember, $0.10 / $0.50. Meta says it ties Claude Haiku 5.5 on the AA index (36) at about a third of the cost, but it trails Haiku on FrontierCode 2.0 (27.2% vs 34.9%).
Decision model: e.g. Strands Decider (see below). Closed-choice tasks only.
The decision-model tier is the surprise. Amazon engineers swapped Claude Sonnet 4.6 for Strands Decider, a Gemma 4 E4B build that returns a probability per option, on 20-category ticket labeling. Latency fell from 4,215 ms to 897 ms, and cost per million tickets went from about $22,200 to about $500. Their gotcha: JSON key order moved output probabilities by up to 0.19, so they sort keys. It’s not for tasks needing explanations, fast-changing label sets or long context, they say. Hightouch found that cheap decision models plus an Opus fallback on low-confidence cases matched Opus accuracy at more than 4x lower cost. Before the fallback, the decision models ran about 6 points behind Opus on hard cases.
All prices and benchmarks above are the publishers’ own, except Artificial Analysis’s numbers.
The play: Put your frontier model on planning and review only, send tool calls and extraction to a cheap worker, and move every yes/no or labeling step to a decision model with a low-confidence fallback. Dual-run each swap on real traffic before you switch. Anthropic · Artificial Analysis · Google · Meta · Strands/Amazon · Hightouch · Hacker News on Fable
3. Your agent’s settings file is a secrets file
Security vendor Apiiro searched public GitHub and found 104,185 copies of Claude Code’s .claude/settings.local.json. 1,122 of them held what it classes as a real secret, including 379 database passwords. Apiiro didn’t test whether any still work, but one company confirmed its exposed key had been live. The mechanism: clicking “Don’t Ask Again” saves the approved command, secret included, to local settings, and the leak happens when that file gets committed. A Claude Code team member replied on Hacker News that the file is meant to stay local and is gitignored automatically when Claude creates it, and that the team is “looking into what we can do to prevent this.” Apiiro’s catch: that ignore rule lives in your machine’s global gitignore, so it doesn’t cover clones on other machines or files already committed, and deleting the file now leaves it in git history. Gitleaks and TruffleHog both missed 215 of the 1,122, because a password after PGPASSWORD= has no key format to match. Days earlier, GitHub issue #107757 (Nov 5) showed Claude Code putting the user’s email address in the User-Agent of web fetches. Anthropic says a default-on privacy mod in v2.1.317 and later blocks it.
The play: Search your repos and their history for .claude/settings*.json, rotate anything you find, add .claude/settings.local.json to your global gitignore, stop passing secrets inline to agent-run commands, and update Claude Code past 2.1.317. Apiiro · The Hacker News · Hacker News on the secrets · GitHub issue #107757 · Hacker News on the User-Agent
Sponsored
Digital Fortress builds AI-agent security packs for teams that put agents in production.
Free: a baseline set of guardrails for coding and ops agents → github.com/digitalfortressaudit/agent-guardrails-baseline
Paid ($99): the Agent Runtime Containment Pack, for keeping agents, and the secrets they touch, inside the lines → fortressaudit.gumroad.com/l/ngouvf
Tool of the Week: Modern Web Guidance 1.0
Built by the Chrome, Edge, Firefox and MDN teams, Modern Web Guidance is an open-source agent skill plus CLI that steers coding agents toward modern web platform features instead of legacy JavaScript and polyfills. Install it with npx modern-web-guidance@latest install, then run legacy-scan . to flag legacy patterns in an existing codebase. It’s free and cross-browser. The play: run legacy-scan on your main frontend repo this week, and install the skill wherever your agents write UI code. GitHub · Hacker News discussion
Check Your Plumbing
GitHub Actions blocks
pull_request_targeton public repos from Dec 2 unless an Actions event policy allows it. The next brownout is Nov 16, 16:00–18:00 UTC. Private and internal repos and GHES aren’t affected. Check Policy Insights, switch topull_requestwhere secrets aren’t needed, or allowlist specific workflows. GitHub Changelog“Sshole” on Alpine/musl: with sshd running without PAM (Alpine’s default), accounts whose shadow password is
*accept any password between 257 and 1,024 characters, and the defaultalpineuser on official AWS/Azure images is in that state, with passwordlessdoas. glibc, PAM, Dropbear and OpenBSD aren’t affected. Upgrade to OpenSSH 10.6p2 or setPasswordAuthentication no, then check logs for password logins on locked accounts. sshhole.com · OpenSSH 10.6p2 notesVal Town is winding down: read-only Dec 10 (excluding SQLite and blob storage), everything else off Jan 10, 2027. Migrate scripts and crons now; it has guides for Cloudflare, Vercel, Netlify and Render. Val Town
Conductor Cloud is joining Meta: cloud workspaces go read-only Nov 13 and are deleted starting Nov 20. Export from Settings this week. Conductor
On the Radar
Inkling 2 Small (Thinking Machines, 594B MoE / 17B active): Apache 2.0 weights were promised for “next week” at the Nov 5 launch, after external safety testing. As of Wednesday evening it was still closed weights on Tinker only, with no Hugging Face repo. Don’t plan around it until the weights actually land. Announcement
Reality check on AI bug hunting: vuln-intel vendor VulnCheck counts 442 CVEs attributable to Anthropic’s Project Glasswing this year, about 0.5% of 93,450 published, with 2 known exploited in the wild. Citing Anthropic’s own Nov 2 ledger, it counts 37,966 findings, 7,349 made public and 217 fixed, and notes curl marked 19 of the first 34 reports from Anthropic’s scanner as not vulnerabilities. VulnCheck sells vulnerability intelligence, so weigh it as a vendor’s read. VulnCheck
That’s the week. Forward it to whoever owns your DNS. Reply and tell us which tier your agents actually run on, or what turned up when you searched for settings files.
The Leverlark team


