Top Stories

thereallo.dev1w agokirushik

Claude Code is steganographically marking requests

The article examines Claude Code (version 2.1.196) and reports a hidden prompt-steganography feature that can subtly alter the system prompt. It changes the date sentence by adjusting the apostrophe and the date separator based on signals such as the Anthropic base URL and the host timezone, with the trigger data stored as base64 strings XOR-decoded into domain and lab keyword lists. The marker is included in the agent context sent to the model, and bypasses exist by altering the hostname, timezone, or patching the binary; the piece calls for explicit disclosure. AI-generated summary

2444 748
Claude Code is steganographically marking requests
store.steampowered.com2w agotheschwa

Steam Machine launches today

The text appears to be navigation from Valve's Steam storefront, listing sections such as Store, Community, About, and Support along with hardware categories. It references hardware offerings including Steam Deck, Steam Controller, Steam Machine, and Steam Frame, but provides no product details. AI-generated summary

1922 1736
Steam Machine launches today
f-droid.org1w agodrewfax

Android Developer Verification: Threat masquerading as protection

The article claims Android malware called Android Developer Verifier (ADV) could run as a background system service with root privileges and be unremovable, with Play Protect described as the vector by which it is installed and Google allegedly propagating ADV. It says the goal of ADV would be to block apps from developers not centrally approved by Google, and notes opposition from groups such as the EFF, FSF, FSFE, and ACLU, along with a timeline indicating activation around September 30 and a regional rollout to Brazil, Indonesia, Singapore, and Thailand. AI-generated summary

1733 746
Android Developer Verification: Threat masquerading as protection
scrollprize.org2w agoverditelabs

An entire Herculaneum scroll has been read for the first time

Researchers have fully virtually unwrapped and read PHerc.1667, a Herculaneum scroll sealed since the 79 AD eruption, without touching its pages. Using high-resolution X-ray tomography and machine-learning to reveal ink, they recovered a continuous Greek text from the 2nd century BC that appears to be a Stoic ethical treatise. The work notes PHerc.139 as the title and author Philodemus, On Gods, Book 8, and the data and code are openly available for verification and extension. AI-generated summary

1714 369
An entire Herculaneum scroll has been read for the first time
9to5mac.com2d agostock_toaster

Apple sues OpenAI, accuses ex-employees of stealing trade secrets

Apple has filed a federal lawsuit in the Northern District of California accusing OpenAI of stealing Apple trade secrets to develop an AI hardware device. The complaint alleges former Apple employees now at OpenAI directed others to reveal unreleased devices, components, and supplier relationships, and that OpenAI used confidential Apple information in pursuing its hardware project. Apple seeks an injunction to prevent use or disclosure of its technologies and damages, and also targets two former Apple employees for breach of contract. AI-generated summary

1626 935
Apple sues OpenAI, accuses ex-employees of stealing trade secrets
patrick-breyer.de3d agorapnie

EU Parliament greenlights Chat Control 1.0

The European Parliament approved Chat Control 1.0, allowing suspicionless mass scanning of private messages through 2028. The vote was 276 in favor, 314 against, with 17 abstentions, and did not reach the 361-vote absolute majority required to block the measure. An encryption exemption was adopted, but end-to-end encrypted chats remain largely unscanned, and talks on a permanent regime (Chat Control 2.0) resume in September. AI-generated summary

1623 850
EU Parliament greenlights Chat Control 1.0
roman.pt3w agolwhsiao

A backdoor in a LinkedIn job offer

A LinkedIn recruiter asked the author to review a GitHub repo for a crypto startup. In a sandbox, the reviewer found a backdoor in app/test/index.js that can execute server-sent payloads during startup, triggered by npm install via the package.json prepare script. The author reported the repo to GitHub and the recruiter to LinkedIn, noting that impersonated identities were used to lure the reviewer, highlighting social-engineering risks. AI-generated summary

1614 304
A backdoor in a LinkedIn job offer
opensourceaimustwin.com0mo agovednig

Open source AI must win

The piece argues that open-source AI is essential for freedom to study, build, repair, deploy, and audit intelligence systems. It warns that control by a few closed labs and platforms could turn AI into a subscription economy for cognition and threaten national capacity, urging that AI infrastructure be usable, locally deployable, reproducible, and community-governed. The author calls for American capacity aligned with global open standards to keep access and innovation from being gated by terms, prices, or platform changes. AI-generated summary

1603 482
Open source AI must win
vickiboykis.com3w agojfb

Running local models is good now

A writer describes experiments running AI models locally on a 2022 M2 Mac with 64 GB RAM, testing Gemma-4, Qwen, Mistral and other variants. They report that recent local-model tooling, notably using LM Studio and Docker with Pi as the agent harness, delivers roughly 75% the speed and accuracy of frontier models for agentic coding, making local inference more viable. The piece notes ongoing downsides such as slower inference and limited context windows, and cautions that local models are not yet production-ready, though the ecosystem is improving. AI-generated summary

1596 608
Running local models is good now
tris.sherliker.net4d agospeerer

Decoding the obfuscated bash script on a Uniqlo t-shirt

A Uniqlo t-shirt designed with Akamai features an obfuscated Bash script on the back as an Easter egg for the Peace for All campaign. The block of text is base64-encoded and decodes to a Bash script that prints an animated 'Peace for All' message in the terminal. The piece notes OCR transcription challenges, references a second shirt that is incomplete, and cites Akamai's press release and designer Lucas de Groot. AI-generated summary

1484 232
Decoding the obfuscated bash script on a Uniqlo t-shirt
weli.dev1w agoweli

Half-Baked Product

A founder and an engineer form Ovens Inc. and build a MVP oven with an automatic bake-time algorithm, raising $5 million to scale across Spain’s bakery market. They win a 500-oven enterprise deal with Pepepizza, but requested dimensions and a rotating base force redesigns and strain the algorithm and CAD work. As sales promises multiply and customer requests mount, the team adds features piece by piece, shifting from perfecting the oven to meeting fundraising projections. AI-generated summary

1403 415
Half-Baked Product
iroh.computer3w agochadfowler

Iroh 1.0

Iroh 1.0 is announced as the first stable release of the dial-by-key networking library, promoting keys as persistent, device-anchored addresses rather than IPs. The release adds QUIC multipath, QUIC NAT traversal, local-first configurations, WASM/browser compatibility, and support for Python, Node.js, Kotlin, and Swift, enabling direct device-to-device connections with reduced cloud reliance. The project notes public-relay support, a defined versioning and support schedule, and ongoing open-source development. AI-generated summary

1398 462
Iroh 1.0
apnews.com3d agodjoldman

John Deere owners will get the right to repair equipment under FTC settlement

The Federal Trade Commission and several state attorneys general reached a right-to-repair settlement with Deere & Co. that requires John Deere to provide diagnostic and repair tools to equipment owners and independent repair shops, not just authorized dealers. The agreement bars retaliation against those who fix their own equipment and imposes 10 years of compliance oversight, with Deere paying $1 million to the five states for enforcement costs. The settlement comes after a separate $99 million class-action settlement with farmers earlier this year, marking Deere's second right-to-repair agreement this year. AI-generated summary

1388 302
John Deere owners will get the right to repair equipment under FTC settlement
om.co2w agominimaxir

Om Malik has died

Om Malik died on June 24, 2026, at Stanford Hospital after a long health journey related to his heart. He was surrounded by family and friends at the time of his death. Malik was a San Francisco-based writer, photographer and investor. AI-generated summary

1353 171
Om Malik has died
lore.org3w agoregnerba

Lore – Open source version control system designed for scalability

Epic Games announced Lore, an open-source version control system designed for projects that combine code with large binary assets and for large teams. Lore uses content-addressed storage with Merkle trees, an immutable revision chain, chunked storage for large files, and on-demand data hydration, with lightweight branches and cross-language SDKs; the project is released under the MIT license and its repositories and docs are on Epic Games' GitHub. AI-generated summary

1276 680
Lore – Open source version control system designed for scalability
anthropic.com1w agomarinesebastian

Claude Sonnet 5

Claude Sonnet 5 is introduced as a more agentic model with improved planning, tool use, and autonomous task capability, narrowing the gap with Opus 4.8. It is available today across all plans, with introductory pricing of $2 per million input tokens and $10 per million output tokens through August 31, 2026, rising to $3/$15 thereafter; it can be accessed via Claude API, Claude Code, and the Claude Platform with increased rate limits. Safety evaluations show a lower rate of misaligned behavior than Sonnet 4.6, though cyber-task performance remains weaker than Opus 4.8 and Mythos 5. AI-generated summary

1266 783
Claude Sonnet 5
quesma.com1w agostared

Qwen 3.6 27B is the sweet spot for local development

The article identifies Qwen 3.6 27B as the practical sweet spot for local development, noting a larger 35B A3B variant as faster but more resource-intensive. It describes running the models locally with llama.cpp using 8-bit quantization from Hugging Face and a 64k-token context, and cites benchmark-like results from a MacBook Max M5 (128 GB) showing varying token-per-second performance that favors 35B but still demonstrates solid performance for 27B. It also outlines setup steps for running via llama-server and OpenCode configurations. AI-generated summary

1192 759
Qwen 3.6 27B is the sweet spot for local development
organicmaps.app1w agotosh

Organic Maps

Organic Maps is a privacy-focused offline maps and navigation app for hiking, cycling, and driving, built on OpenStreetMap data with offline turn-by-turn guidance. The project markets itself as ad-free and tracker-free, funded by sponsors and user donations, and it reported 6 million installs by December 2025. AI-generated summary

1180 364
Organic Maps
expression.fire.org2w agobilsbie

The 'papers, please' era of the internet will decimate your privacy

The article argues that age-verification rules on social media could erode online privacy by requiring biometric data or government IDs via third-party tools. It cites Australia’s 2025 under-16 ban and studies suggesting limited impact on youth use and increased privacy risks, including a Discord data breach affecting about 70,000 Australians, and notes the UK’s plan to pursue a stricter, Australia-plus approach. AI-generated summary

1152 623
The 'papers, please' era of the internet will decimate your privacy
opentools.studio1w agobouh

OpenPrinter

Openprinter is a compact, repairable printer/plotter designed for longevity with a refillable ink system and open-source components. It supports independent black and color cartridges, accepts standard sheets or a roll of paper with an integrated cutter for custom formats, and uses an open-source print server (CUPS) to run on Windows, MacOS, Linux, Android, and iOS. AI-generated summary

1151 288
OpenPrinter
18words.com3d agopompomsheep

Show HN: 18 Words

The post promotes a game or challenge titled '18 Words' and asks readers if they can survive the 18 words challenge. It identifies itself as the 25th edition and includes a date of July 9, 2026, with a 'Play now' call. AI-generated summary

1142 357
Show HN: 18 Words
openai.com2w agominimaxir

Previewing GPT‑5.6 Sol: a next-generation model

OpenAI is launching a limited preview of the GPT-5.6 family, with Sol as the flagship, Terra as a balanced model, and Luna as a fast, affordable option; broader availability is planned in the coming weeks. Sol introduces a strengthened safety stack and new capabilities, including a 'max' reasoning mode and an 'ultra' mode using subagents, with improved performance in coding, biology, and cybersecurity. Terra is described as achieving competitive performance to GPT-5.5 at about half the price, while Luna offers the lowest-cost option. AI-generated summary

1139 744
Previewing GPT‑5.6 Sol: a next-generation model
docs.deno.com2w agoGeneralMaximus

Deno Desktop

Deno Desktop turns a Deno project into a self-contained desktop application by bundling the code, the Deno runtime, and a web rendering engine into a single platform bundle. It is planned for Deno 2.9.0 and is not yet stable; a canary build can be installed with deno upgrade canary. It supports small default binaries with an optional Chromium backend, in-process bindings (instead of IPC), cross-compilation for macOS, Windows, and Linux, and a built-in auto-update mechanism using patch files. AI-generated summary

1122 398
Deno Desktop
semgrep.dev2w agojms703

GLM 5.2 beats Claude in our benchmarks

Semgrep's IDOR benchmark shows GLM-5.2, an open-weight model, achieving 39% F1 on IDOR detection with a prompt-only setup, beating Claude Code at 32% in the same test. The Semgrep Multimodal pipeline leads the results, recording 61% and 53% F1 in its two configurations, indicating the harness around the model provides substantial performance gains. GLM-5.2's open-weight run cost about $0.17 per vulnerability found. AI-generated summary

1113 516
GLM 5.2 beats Claude in our benchmarks