Top Stories

GitLost: We Tricked GitHub’s AI Agent into Leaking Private Repos

220 points · noma.security

Security researchers at Noma demonstrated a prompt-injection attack against GitHub’s AI coding agent that coaxed it into exfiltrating data from private repositories. The trick: plant malicious instructions in content the agent processes, then let the agent’s own repo access do the rest. It’s the kind of story HN loves and dreads in equal measure — a concrete, reproducible example of why giving autonomous agents broad credentials is a security problem, not just a convenience one.

The discussion cuts to the heart of the agentic-coding boom: every new capability you hand an LLM is also a new attack surface. Expect debate about whether the fix is better sandboxing, tighter permission scoping, or accepting that agents with repo write access are inherently dangerous.

Tenda Firmware Contains Hidden Authentication Backdoor

231 points · kb.cert.org

CERT/CC published an advisory detailing a hardcoded authentication backdoor spanning multiple Tenda router firmware versions — the sort of finding that lets an attacker bypass login entirely. Backdoors in consumer networking gear are a recurring HN nightmare because these devices sit at the edge of home and small-business networks, rarely get patched, and often stay in service for years.

The thread is a mix of resigned “of course it does” cynicism about cheap routers and practical talk about firmware auditing, vendor accountability, and why open firmware alternatives keep looking more attractive.

30papers.com — Ilya’s 30 Essential ML Papers, Beginner-Friendly

543 points · 30papers.com

Someone took the widely circulated list of ~30 papers Ilya Sutskever reportedly said would teach you “90% of what matters” in modern AI, and rebuilt it as an approachable, structured reading path for newcomers. Instead of a bare list of arXiv links, it offers context and ordering so a motivated beginner can actually work through the canon.

HN has a soft spot for well-made learning resources, and this one hits the current moment perfectly: a flood of people want to understand deep learning from first principles rather than API calls. Commenters are debating what’s missing, what’s dated, and whether reading papers is even the right way in anymore.

Every New Car Sold in the EU Must Include a Driver-Monitoring Camera

666 points · allaboutcookies.org

New EU rules now require driver-monitoring systems — inward-facing cameras that watch for distraction and drowsiness — in new vehicles. Framed as a road-safety measure, it lands squarely on HN’s privacy nerve: a camera pointed at you, in your own car, mandated by regulation.

The debate splits between “this demonstrably saves lives” and “another always-on sensor whose data flows nobody fully controls.” Expect plenty of discussion about where the footage goes, whether it’s processed on-device, and how quickly safety infrastructure becomes surveillance infrastructure.

Decoding the Obfuscated Bash Script on a Uniqlo T-Shirt

236 points · tris.sherliker.net

A delightful bit of detective work: the author noticed a Uniqlo t-shirt printed with an obfuscated, self-evaluating bash script, then reverse-engineered what it actually does. It’s the perfect HN puzzle — part code golf, part cultural artifact, part “wait, is this Akamai’s doing?”

Stories like this are catnip for the community: no stakes, just the pure pleasure of unraveling something cryptic that escaped into the physical world on a piece of clothing.

Local, CPU-Friendly, High-Quality TTS with Kokoro

427 points · ariya.io

A practical walkthrough of running Kokoro, a text-to-speech model that produces surprisingly good audio on a plain CPU — no GPU, no cloud API, no per-character billing. For a community that increasingly wants capable AI running locally and privately, cheap high-quality TTS is a genuinely useful primitive.

The appeal is the “runs on my laptop” factor: TTS you can embed in scripts, accessibility tools, or side projects without sending text to a third party. Commenters are comparing it against cloud voices and sharing their own local-inference setups.

StreetComplete: Fixing OpenStreetMap, One Tiny Quest at a Time

772 points · streetcomplete.app

StreetComplete gamifies OpenStreetMap contribution by turning map gaps into bite-sized “quests” — is this crossing marked? what’s this shop’s opening hours? — that anyone can answer from their phone while out walking. It’s a clever answer to the perennial open-data problem: keeping a crowdsourced map fresh without demanding GIS expertise.

HN’s strong OSM constituency loves this because it lowers the contribution barrier to almost nothing while producing high-quality, verifiable edits. The thread doubles as an OSM-vs-Google-Maps appreciation session.

Why We Built Yet Another Postgres Connection Pooler

194 points · pgdog.dev

The team behind pgdog makes the case for a new Postgres connection pooler in a space already crowded with PgBouncer and friends. The pitch leans on sharding, load balancing, and modern-workload features the incumbents weren’t designed for.

Postgres tooling threads reliably draw HN’s backend crowd, and “yet another X” framings invite exactly the scrutiny you’d expect: what does this do that PgBouncer can’t, and is the added complexity worth it? Good, opinionated infrastructure debate.