Top of Hacker News

LiteLLM 1.107.0 reached PyPI at 11:39 UTC on 21 October and ran a credential stealer inside every Python process on the host. A hidden daemon then phoned home hourly for more code, and by 13:40 the release was gone.

Would a dependency cooldown have caught it? One supply-chain engineer in the thread says even 24 hours is enough. A skeptic says the median attack stays hidden for months.

Countries' current pledges would cut emissions 5 to 8% by 2030, against the 18% UNEP says is needed. Getting back under 1.5°C also means pulling about 320 gigatons of CO2 out of the air by 2100, roughly eight times what the world emitted in 2024.

A climate-tech worker with 12 years in the field answered with a plea for stamina: the physics never changes, so neither does the task.

Claude Duet is a research preview for US Pro and Max subscribers, powered by Opus 5.5. It runs larger tasks, such as scheduling meetings or building dashboards, on its own cloud computer, and reports back when done.

The first complaint in the thread: a lab with frontier coding agents could let you ask Claude for a spot instead of retyping your details.

PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35 shipped on 21 October, a day ahead of plan, after the CVE IDs leaked. Both bugs score CVSS 9.2 and need no login, and a proof-of-concept video went up by mistake.

In the thread, an admin racing across several servers said a three-line nginx rule blocking multipart uploads had been quietly protecting a machine he could not shut down.

The release adds a Quack server protocol, background reads for Parquet and CSV, VARIANT with automatic shredding, and triggers. A new PEG parser is now the default, ICU is gone, and 2.0 is supported until at least October 2027.

One user reports shrinking 7 TB of Azure NDJSON logs to 2.5% of their size in Parquet in under two hours on a medium VM.

Changing only the customer's name in the prompt moved the share of hackable apps, and for a Hong Kong democracy group nearly half the apps leaked user secrets, against about a third for a neutral customer. The authors call it a backdoor.

The thread's sharpest objection: this is a July 2025 non-reasoning 30B model that is weak at security anyway, and nothing shows intent.

Railway believes the attack is part of a wider campaign against platform companies and AI inference providers, and has asked upstream providers to revoke keys. Deployments are paused during the investigation.

In the Railway Discord, users say stolen keys were already being used, and one says usage climbed on keys created only for Railway, 10 hours into the incident.

Practical Engineering traces the fire escape from a cost shortcut to a ban for new construction in 1968, with the 1911 Triangle Shirtwaist fire in between, where an exterior escape buckled under the crowd.

Commenters argue over whether to miss them: one says everything else about fire safety got better, another would still risk a rusty escape over a blocked stair.

An essay at hrih.si says AI follows FORTRAN's pattern, only faster: a METR trial found experienced developers 19% slower with AI while believing they were faster, and Faros AI data showed more pull requests but much longer reviews.

Commenters attack the analogy itself. One says compilers were first called "automatic coders"; another says a compiler maps input to output roughly 1:1, and an LLM does not.