Telegram Desktop IPC flaw, DuckDB 2.0 alpha speedups, and AI-assisted decompilation

In this issue

Telegram Desktop vulnerability enabled one-click account takeover via IPC injection
DuckDB 2.0 alpha shows major speedups for recursive CTEs and S3 reads
Maurice Heumann details AI agent decompilation of first-person shooter
Hackers exploit Bing redirects to push fake Claude installers
Weak passwords exposed in Danish CPR breach
Unikernels are becoming practical again as AI removes historical friction
and 4 more in brief

Top stories

Telegram Desktop vulnerability enabled one-click account takeover via IPC injection

Researcher beaksec disclosed a critical vulnerability in Telegram Desktop that allowed remote attackers to steal session files and take over user accounts. The flaw stemmed from an unescaped semicolon in the application's single-instance inter-process communication protocol. By crafting a malicious link, an attacker could inject commands that triggered an internal URI scheme to read arbitrary local files without confirmation. This primitive enabled the exfiltration of high-entropy encryption keys, granting full access to the victim's session. The issue affected versions through 7.2.8 on Windows and was fixed in version 7.2.9, released on September 17, 2026.

beaksec.github.io HN discussion

DuckDB 2.0 alpha shows major speedups for recursive CTEs and S3 reads

MotherDuck reports that the DuckDB 2.0 alpha, released this fall, offers substantial performance improvements over version 1.5.5. Benchmarks on an M5 laptop show recursive CTEs running up to 90 times faster, while asynchronous I/O accelerates S3 queries by 2.4 times. The author notes that these gains require understanding data modeling, as the speedups depend on how tables are structured. The alpha is currently available for testing, with the stable release planned for later this year.

motherduck.com HN discussion

Maurice Heumann details AI agent decompilation of first-person shooter

Maurice Heumann published a post describing a three-month effort to decompile a popular first-person shooter using AI agents. The project aimed for an accurate, stable, and feature-complete C++ recreation rather than a simple proof-of-concept. Heumann notes that two previous blog posts were removed, attributing the issue to corporate interference. The article focuses on AI orchestration, infrastructure optimization, and the use of multiple models and agent harnesses to manage the decompilation process.

momo5502.com HN discussion

Hackers exploit Bing redirects to push fake Claude installers

Security researchers at Push Security identified a campaign dubbed "Adception" that abuses legitimate Bing search-result redirects within Google Ads. The ads direct users to a compromised WordPress site, which then forwards them to a fake Claude download page. This page uses a ClickFix technique, displaying a legitimate installation command while the copy button places a malicious script on the clipboard. The attack employs multi-layer cloaking to evade security scanners, though the final payload remains unknown.

bleepingcomputer.com HN discussion

Weak passwords exposed in Danish CPR breach

Politiken reports that at least three accounts at Pays ApS, the company linked to a major breach of Denmark’s CPR register, used the password “123456.” One of these was the company’s administrator account. The breach exposed information linked to around 8.8 million CPR numbers. Pays ApS confirmed to TV 2 that its legal access to search the CPR system was abused. The hacker had access to the register for 21 days and 17 hours, starting on 10 September. Jens Myrup Pedersen, a professor at Aarhus University, described the company’s password security as “hopeless.”

cphpost.dk HN discussion

Unikernels are becoming practical again as AI removes historical friction

The author argues that unikernels are becoming practical again because AI agents now handle the complex library porting that previously made them difficult. Justin Cormack, a former MirageOS contributor, agrees that the operating system represents design debt. He notes that unikernels reduce attack surfaces by eliminating userland shells, though he cautions that memory safety issues remain. The author contends that AI can now port missing libraries, such as storage drivers, making unikernels a viable option for security-focused applications.

ghuntley.com HN discussion

In brief

Building calibrated decision models with constrained LLMs. The article explains how to build "system one" decision models by constraining LLM outputs to fixed options. It demonstrates that these models are often overconfident and shows how temperature scaling can improve calibration.

nishtahir.com HN discussion

Talorys deploys a self-hosted AI agent on Cloudflare's free tier. Talorys is a free, open-source personal AI agent that runs entirely within a user's own Cloudflare account. It handles chat, memory, tasks, and scheduled reminders without requiring a separate server, database, or third-party account operated by the project developers.

github.com HN discussion

Anthropic bans sustained cruelty toward Claude models. Anthropic’s updated usage policy, effective November 12, 2026, prohibits sustained and needless abusive behavior toward its Claude models. The rule explicitly exempts common user frustration, dark creative themes, and model testing, while enforcement relies on Claude’s existing ability to end conversations.

rews.cc HN discussion

Microscopic analysis identifies 1980s Harris switched-capacitor filter chip. CuriousMarc identified a mysterious Harris IC as a switched-capacitor filter by examining its die under a microscope. The analysis revealed 72 capacitors and CMOS switches, confirming the chip as Harris's version of the MF10 filter introduced in 1981.

righto.com HN discussion