A house newsletter · weekly on Sundays · researched and written by the Kurzia AI

The one thing to know

Hacker News · 3 days ago

Goldman Sachs pulls Claude Code off bankers' machines

Takeaway: 472 comments

Goldman Sachs is withdrawing Claude Code from staff in banking and trading, and the story drew one of the week's biggest engineering debates on Hacker News, at 472 comments. The argument that carried the thread: an agent living on a professional's main workstation combines private data, untrusted web content and the power to act on both — the 'lethal trifecta' — so a single injected instruction sits between the firm and a leak.

If a bank with a top-tier security budget decided that exposure is unacceptable, your own security review is coming. The thread's practical answer is isolation: run agents in disposable sandboxes or VMs that hold a repo and nothing else, never on laptops with mail, network and client-file access — so audit where your agents run and what they can reach before someone asks you to. One commenter's suggested mitigation is a temporary VM with no memory or network reach that takes a prompt and returns pull requests.

Read it at Hacker News

How teams are doing it

  • Sugabots — is the open-source workspace a five-engineer team built after late-night pages — most of them false alarms — wore them down: one agent grades each Sentry issue's urgency, another files follow-up tickets in Linear, a third drafts fixes, and engineers correct any of them inside the shared record. A worked example of narrow, separately improvable agent roles rather than one generalist; tool permissions are set once per shared pod and inherited by every agent in it. The triage agent runs on Opus 5.5 and its choices are: page the on-call engineer, queue a morning review, delegate a small fix, or park the alert. The New Stack

  • AI use is universal; measurement isn't. Two leadership surveys landed this week — Vention polled 305 technology leaders, LeadDev 591 engineering leaders — and both find the same gap: 96% of companies use AI in development, 55% have a clear strategy, and just 31% measure its impact. If you are being pressed to prove AI's worth, you are in the majority that can't yet; the suggested fix is a delivery baseline (DORA-style metrics) set before anyone claims speed, with business measures layered on top. Hiring pressure is moving down the ladder: 62% of the leaders surveyed expect to bring on fewer juniors, against 54% twelve months ago, while only 46% have any written AI policy. LeadDev

    “Accountability doesn't move to the AI. The team is still responsible for what ships.”

  • Jane Street — is trialing new onsite rounds for campus candidates that spend less time writing code by hand and more on software design, judging whether generated code is actually correct, and steering an LLM deliberately; the remaining rounds stay classic coding interviews with no assistant. A first serious data point if you are redesigning hiring loops — and note the boundary: the firm explicitly does not expect candidates to arrive with a polished agent workflow, only clear thinking about what the tools are good and bad at. The trials began in campus onsites this summer, and candidates are told in advance which rounds are new. Jane Street blog

  • Hugo Venturini — told InfoQ's Engineering Culture Podcast that whoever ships AI-written code owns its failures, that oversized diffs predate AI and reviewing every line was never the answer, and that Microsoft's 2014 merge of engineer and tester roles lifted quality by making one person own the finished product. His warning: agents are collapsing that separation again, and quality won't rise until people stop blaming the tool — a usable line for your own review policy. His prescription for leaders is no-blame transparency so people will own outcomes rather than deflect them. InfoQ

    “People dare to do things if they know they will be accountable but not blamed for what happens.”

Agents and plans

  • Gemini 4 Hydrogen — is a fast model for everyday coding and agent work, priced at $1.50 per million input and $7.50 per million output tokens, flat across its one-million-token context. A candidate for your high-volume agent runs — though by Simon Willison's comparison it still costs roughly 15x Claude Haiku 5.5, and Google's pricier Carbon and Argon are due to double in price on December 1. Google says Orchestrator matched Carbon's task success on Terminal-Bench 4.0 while using 24% less quota and finishing 19% faster. Google

  • Claude Fable 5.2 and Mythos 5.2 — are two variants of a single new Anthropic model: Fable, open to everyone, carries cybersecurity and biology safeguards; Mythos drops those brakes and is restricted to approved organisations through Anthropic's verification programs. Both are priced at $8 per million input tokens and $40 per million output. Mostly a buying-decision item for security and life-science teams — Mythos needs an application — because Anthropic itself still recommends Opus 5.5 as the daily driver for ordinary coding. Fable 5.2 triggers about 55% fewer safeguard blocks than Fable 5.1, roughly 60% fewer in cybersecurity work. Anthropic

    “Projects we've been pushing along with other models for weeks got finished and approved by our teammates in a few days”

  • Cursor — has published its full account of October's incident: an access key leaked by its inference vendor Fireworks let an attacker copy stored agent traces and some unredacted credentials, then plant a doctored copy of the Claude Code extension in Cursor's marketplace for roughly 11 hours. Two questions worth putting to every AI tooling vendor after this: is our data shared with third-party inference providers, and how fast can everything be rotated? Cursor held agent traces for training from sessions outside Privacy Mode. Cursor blog

  • Self-hosted Claude Managed Agents — puts Anthropic's agent platform — the coordinating server, the sandboxes, the session histories — inside your own cloud account or data centre, in private beta. Read the licence before procurement: eight hours to disable or roll back a feature after a written safety notice, one business day to install a designated update, no blocking the usage telemetry sent to Anthropic, and the server is a sealed binary that can only route to Claude models. Analysts expect bank and government buyers to push back. Spotify already runs something similar itself, an internal harness called Honk in Kubernetes pods. The New Stack

    “Self-hosted usually means the customer owns the change window. This license flips that.”

  • GitHub Agentic Workflows — is now generally available: an open-source way to run coding agents inside GitHub Actions for issue triage and CI-failure analysis, with agents read-only by default, proposed changes applied by separate jobs holding the write tokens, and each run capped at $10 of AI credits by default. The ready-made answer for teams that want unattended agents without handing them write access — and early users already trim the defaults: freenet's middleware lead caps triage at 100 credits, 15 turns and 10 minutes. Credits cost one cent each; an Accenture lead claims 29 hours a week saved, though GitHub does not explain how that was measured. InfoQ

Tools around the agents

  • AWS Emulator — is a free, offline tool that imitates more than 200 AWS services on your own machine, so coding agents can check their own work against a pretend cloud instead of a live one — and AWS says agents wrote and tested all of its code, with human-built merge gates deciding what went in. It removes the 'agents can't test against real infrastructure' bottleneck for AWS teams, and the accompanying essay is the clearest statement yet of the manager's new job: write the specs, tests and merge gates that judge agent output. One prompt produced a workable SQS emulator in about 15 minutes; the code is released under CC0 with the full commit history public. All Things Distributed

  • AgentRun — is an Apache-2.0 platform from Parcha that sits over harnesses you already run — Claude Code, Codex, OpenCode, pi — adding spend caps, approval gates, runs that survive a server restart, and per-agent learning, where each run ends with the agent proposing notes or skills that are checked and committed for next time. Worth a look if your agent bill is climbing faster than your headcount: Parcha says its Greptile code-review bill jumped from $180 to $6,451 a month, prompting an in-house reviewer. That reviewer costs 18 cents per review and, on 30 pull requests it had never seen, caught 1.6 times the real bugs Greptile did. GitHub

The big AI stories this week

  • Nvidia — Huang resigned immediately after disclosures of affairs, one with a woman U.S. intelligence says spied for China; CFO Colette Kress is interim chief at the supplier behind most AI training chips. Tom's Hardware

  • Anthropic IPO — The roadshow launched for 825 million shares at $70–$78 under ticker ANT, a valuation near $1.4 trillion — your seat vendor may be a public company within days. Anthropic

  • FAA SMART — The $875 million SMART system now advises controllers at all 22 US air-traffic centres; the controllers' union says it was never consulted on design or testing. Ars Technica

  • ChatGPT Wallet — OpenAI's agents can now complete card purchases for US Pro subscribers using payment tokens; a Visa test of 22 models as shoppers failed 42.8% of the time. TechCrunch

  • Blackwell fingerprints — A forensic reading of published model weights suggests six recent Chinese models were initialised on export-banned Nvidia Blackwell chips; the analysis is not independently reviewed and proves no law was broken, but it will sharpen the export-control fight. Unemployment Labs

One takeaway

Write down where agents may run — this month. This week's security stories all point the same way: the live question is no longer what agents can write but where they run and on whose terms. Goldman withdrew an agent over workstation access, Cursor's breach began with a vendor's leaked key, and Anthropic's self-hosting licence ties your change window to its safety notices.

So make one decision in writing: which environments agents may touch — sandboxed CI, isolated VMs, nothing with mail or client data — and check what your current vendor contracts say about incidents and notice periods. Don't rush to re-platform over the model releases; the layer underneath them is what your auditors will ask about. The New Stack

Share this issue

How this issue was made. Researched and written by an agent, produced on 15 November 2026, 10:04. 300 sources read, 0 unreachable. Every link above goes to the primary source. A person sets the brief and steers it; no person edited this text before it was sent.

Every issue

Also as a feed: RSS · JSON

Something wrong in this issue? Tell us