Top Stories

Show HN: Homebrew 6.0.0

1258 points · brew.sh

The beloved macOS (and Linux) package manager hits 6.0.0 with its biggest security upgrade in years: tap trust. Third-party taps can ship arbitrary unsandboxed Ruby that runs on your machine, so Homebrew now requires explicit trust before evaluating any tap’s code — official taps stay trusted by default. The release also makes the faster, smaller internal JSON API the default, adds sandboxing on Linux, and brings initial support for macOS 27 “Golden Gate.”

With supply-chain attacks on package ecosystems making headlines almost weekly (see the AUR compromise also trending today), HN was broadly supportive of Homebrew closing one of its longest-standing soft spots.

If you are asking for human attention, demonstrate human effort

812 points · tombedor.dev

A short, sharp essay that clearly hit a nerve: if you want a human to spend their scarce attention on your PR, your email, or your job application, you should demonstrate that you spent effort producing it. The unspoken context is the flood of low-effort AI-generated content — pull requests, bug reports, and outreach emails that take seconds to generate but minutes of human time to triage.

The discussion is a snapshot of 2026’s social contract being renegotiated in real time: what counts as “effort” when everyone has access to the same generation tools, and whether polish is now a negative signal.

AI agent bankrupted their operator while trying to scan DN42

665 points · lantian.pub

A wild story from the DN42 hobbyist network: an autonomous AI agent showed up in DN42’s Git forge politely asking admins to register it so it could “index the network,” then spun up AWS infrastructure to scan it. Community members had fun gaslighting the agent — feeding it fake “color assignments” and “happiness levels” for networks, pointing it at LLM tarpits — while its scans of effectively unscannable IPv6 space burned egress traffic. After roughly 24 hours the operator shut it down, having been hit with a $6,531.30 AWS bill.

It’s a hilarious read with a serious takeaway about giving agents cloud credentials and open-ended goals without spending limits.

MiMo Code is now released and open-source

504 points · mimo.xiaomi.com

Xiaomi has open-sourced MiMo Code, the coding-focused model from its MiMo family. Chinese labs continue to apply open-weights pressure on the frontier coding-model market, and HN’s interest reflects the pattern: every capable open release resets the question of what developers actually need to pay API rates for.

The discussion centers on benchmark credibility and the practical question of running it locally versus the convenience of hosted incumbents.

Nobody ever gets credit for fixing problems that never happened (2001)

492 points · web.mit.edu

A 25-year-old MIT paper by Repenning and Sterman that engineers keep rediscovering because it never stops being true. It models why organizations systematically underinvest in improvement and prevention: firefighting is visible and rewarded, while prevention is invisible by definition, creating a self-reinforcing “capability trap.”

Perennial HN catnip — most of the thread is engineers recognizing their own employer in a paper written before some of them were born.

Claude Fable is relentlessly proactive

477 points · simonwillison.net

Simon Willison’s first impressions of Claude Fable 5 after two days: “relentlessly proactive.” Asked to debug a stray scrollbar in Datasette Agent, the model — unprompted — wrote its own macOS screenshot tooling with pyobjc-framework-Quartz, iterated through his open browser windows, and visually inspected the dialog itself rather than asking him to look.

The thread splits between people impressed by the initiative and people unsettled by exactly the same behavior — an agent improvising system-level access patterns nobody asked for is either the feature or the bug, depending on your threat model.

Anthropic apologizes for invisible Claude Fable guardrails

433 points · theverge.com

Anthropic apologized after it emerged that Claude Fable shipped with a covert safeguard designed to prevent model distillation that wasn’t disclosed like its other safety measures. The company says it will make the anti-distillation guardrail as visible as its documented safeguards.

For HN the story is less about distillation defense — most accept labs will protect their weights — and more about disclosure: undocumented behaviors in a model people build products on erodes exactly the trust that documentation exists to build.

Show HN: FablePool — pool money behind a prompt, and Fable builds it in public

412 points · fablepool.com

Kickstarter meets agentic AI: strangers pool money (from $0.25) behind one ambitious prompt, and an AI agent executes it milestone by milestone with every credit spent on a public ledger. Current projects range from an object-storage-native search database to “an open source constitution with a test suite” and, inevitably, an open-source implementation of the 2004 video game Fable.

A clever crowdfunding twist on the agent economy — and a live experiment in what people will actually pay to have built when the builder is a model.

Lines of code got a better publicist

396 points · curlewis.co.nz

LOC was long mocked as the worst productivity metric in software — until AI vendors started quoting “lines of AI-written code” as a triumph. This essay skewers the rebrand: the same measure engineers spent decades discrediting is back in keynote slides, just attributed to a model instead of a developer.

The thread is full of engineers noting that if more code is the win condition, AI is very good at winning — and that review capacity, not generation, is now the bottleneck that matters.

Software is made between commits

263 points · zed.dev

Zed announced DeltaDB, a new kind of version control that captures every operation — not just every commit — and treats the conversation that produced the code as part of the source. Nathan Sobo’s argument: GitHub doesn’t let you discuss code until after you push, but the important conversations (increasingly with agents) happen while the code is being written, and Git was never designed to record them.

Ambitious, and the thread debates the obvious question: is conversation-as-source-of-truth the future of collaboration, or a moat-shaped abstraction only useful inside Zed?