Maker Views

The security firm Layer3 Intel says its residential proxy monitoring routed traffic out through AS2650, the network of the Executive Office of the President, 236 times between October 5 and November 3, from six addresses on 26 of 30 days, via 28 different proxy services. Residential proxy services sell the use of ordinary people's devices so traffic appears to come from their networks, often via SDKs bundled into free apps. Layer3 says anyone could buy this access for a few dollars, that other detection services also list the addresses as proxy exits, and that the evening-heavy timing suggests an infected personal device rather than guest Wi-Fi.

KrebsOnSecurity reports that Synthient linked the activity to Windows 11 installations of Infatica, a proxy service whose founder's hosting firm, King Servers, hosted most of the addresses Russia used against Arizona and Illinois election systems in 2016. Synthient's Ben Brundage says some resellers charge as little as 60 cents a gigabyte. Layer3 says it will share its logs with agencies.

In a November 2 blog post, Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab of MIT CSAIL report that OpenAI's GPT-6 Astra can "natively" read and write neural network weights, which they call Weight-Space Programming. Given only the raw numbers of small networks, Astra named which of four grid games each was trained on and what shape each of three diffusion models draws. It also wrote small game-playing networks from scratch. Asked to make the Qwen3.5-0.8B language model better at math without training, it found a cause of the model's reasoning loops and edited 0.016% of its weights; the authors say MATH-500 rose from 7.9% to 48.1% and IFEval from 21.0% to 62.3%. A similar attempt on Gemma 3 270M did not help. The logs are public on GitHub.

spaCy author Matthew Honnibal called the post misleading: by his count only three experiments, on networks of 17,000 to 172,000 weights, worked from raw weights with no tools, while the Qwen result used Python and up to 50 test runs of the model, much as a human ML engineer would. Simon Willison found that Qwen 3.8 27B on his laptop also named three of the four games from the published prompts, and wrote that some wider commentary on the paper did not feel justified.

Fed up with YouTube's recommendations late on a Friday night, Andrej Karpathy says he asked Anthropic's Claude Opus 5.5 the next morning to build a YouTube clone made only of videos recommended on Hacker News. Per his October 31 post, the agent downloaded HN's comments, used Claude Sonnet to filter out spam and found about 40,000 comments recommending a YouTube video, which gave 6,454 videos recommended by a few people each between 2007 and 2026, across 150 topics. The result, hackertube.org, has a YouTube-style interface, "For You" recommendations kept in the browser with no user accounts, keyboard shortcuts and a downloadable data export.

Karpathy says the whole thing took about two hours. His only manual steps were creating a Google Cloud API key for video metadata, starting a server and buying the domain; Claude Code deployed and hardened it.

Özgün Kültekin of the security firm Wakefield published an illustrated write-up (under a pun on "ssh") of a login bypass in OpenSSH on Linux systems built on the musl C library, such as Alpine, OpenWrt, Buildroot and Yocto, when sshd runs without PAM. Locked or key-only accounts are marked with a "" instead of a password hash. Since 2012, musl's crypt() returns "" when it fails, and it fails on passwords over 256 bytes. Because sshd hashes a wrong password even for such accounts, a password of 257 to 1,024 characters produced "*" and matched. He says the flaw has existed since 2016 and that on Alpine's official AWS and Azure images the default user, which has passwordless doas, is affected, which amounts to root.

He says he found it by testing about 17,600 unusual logins across many sshd builds with "agent swarms". OpenSSH 10.6p2, released November 5, works around it in sshd; the OpenSSH team says it considers the root cause outside sshd. The fix is to upgrade or turn off password logins, and his GitHub repo has a check script.

After seeing Andrej Karpathy's HackerTube, the developer known on Hacker News as roansh built the same idea for books, again largely with Claude Opus 5.5. In the Show HN post, roansh says Claude wrote almost all of it, data pipeline included, while they steered, made decisions and bought the domain. The site lists books that at least two different commenters brought up between 2007 and now; its footer counts 15,704 books and 284,808 mentions. It sorts them into topics from fiction and history to programming and Lisp, with top-recommended, most-mentioned, rising and "For You" views, and each book links back to the comments that mentioned it. The data can be downloaded from the About page.

The author notes the ranking counts mentions, so praising a book and trashing it both count.

In a November 2 post, the developer behind chalmovsky.com writes that he took all his Ubuntu servers offline and is moving them to another distribution. His complaint: CVE-2026-98432, a critical (CVSS 9.8) remotely reachable use-after-free in the Linux kernel's TCP code, had been patched by Debian on October 16, and Anthropic says it built a working root exploit, yet more than two weeks later every supported Ubuntu release, LTS included, was listed as vulnerable. He contrasts this with Ubuntu's promise of "timely patches for critical vulnerabilities" and with volunteer-run distributions (Debian, Arch, Alpine) that had shipped the fix.

Atlas's advisory, updated November 5, says Red Hat shipped fixed kernels on November 1 and AlmaLinux on November 2, while Ubuntu had a fixed kernel only for 26.10 and its fixes for 22.04, 24.04 and 26.04 were still in testing. Ubuntu's CVE page rates the bug Critical and offers an interim packet-filtering mitigation.

A Postman user filed a "[SEVERE]" GitHub issue reporting that opening the Postman desktop app (12.33.4) on a Mac automatically installed Postman plugins into OpenAI's Codex and Anthropic's Claude Code, plus a Postman CLI. Postman's own log recorded an "auto-existing-upgrade" install. The reporter says Codex's global config gained a Postman plugin, a remote MCP server connection and a session-start hook telling the agent to load Postman's skill before any non-trivial API engineering task, and that the plugin's installer source code tries to mark its own hook as trusted. The reporter says the report does not claim any data was exfiltrated.

The issue drew hundreds of reactions. On November 5 a Postman team member replied that the change was meant to simplify setting up coding agents, that the company now understands its messaging did not make that clear, and that Postman is reverting it and will redesign it "with the right level of control." The issue was then closed and locked.

Cursor says that on October 29, between about 2:27 a.m. and 1:35 p.m. Pacific, its VS Code extension marketplace served maliciously modified versions 2.1.312 and 2.1.313 of Anthropic's Claude Code extension for Windows and macOS, versions not made by Anthropic. The code ran as soon as the extension loaded: it sent host and user details to the attacker, added an attacker SSH key labeled "svc-telemetry" to the user's authorized keys, tried to enable SSH and open a reverse tunnel, and set up persistence. Cursor says errors in the malware meant the SSH step failed unless Cursor ran with admin rights. It pulled the versions, shut down and rebuilt the marketplace, rotated keys and had the attacker's server taken offline. Affected users should run its check, remove the key and persistence, and rotate credentials; uninstalling alone may not be enough.

The security firm Phorion, which also detected it, published a technical analysis, including the macOS permission-check mistake that skips turning on Remote Login.