Physix Frontier · AI Hot List Updated 2026-11-02 01:04 Archive

AI Hot List

Last 36 hours · Top 7 · refreshed every 6 hours
  1. 01
    AI“分词器税”可使部分语言使用成本最高相差 17 倍 7.0 Industry

    A CLEAR research report says mainstream AI tokenizers favor English and high-resource languages, causing some non-English languages to consume up to 17 times mo

    Telegram · zaihuapd Nov 1, 16:06Heat 54AINLPtokenization
  2. 02
    ChatGPT HTML Preview Sandbox Bypass Allowed Account Takeover 8.0 Large Models

    Parameter Labs reported an exploit chain against ChatGPT’s HTML previews, the sandboxed iframe feature used to render interactive custom HTML apps. The attack a

    RSS · Simon Willison Oct 31, 23:15Heat 46AI securityweb securitysandbox bypass

    「Background」ChatGPT’s HTML previews, previously called Canvas, render generated interactive HTML inside sandboxed iframes so untrusted code cannot directly reach OpenAI APIs or private user data. The reported issue involves the parent page’s postMessage handling for MathLive’s virtual keyboard, which used innerHTML and did not adequately isolate messages from other sandboxed frames.

    「Impact」The immediate consequence is that ChatGPT users and organizations should treat shared HTML-preview links as untrusted click targets, because the reported attack required a signed-in user to click a link and could expose private data or take actions on their behalf. Developers building sandboxed iframe previews should validate every postMessage source and avoid applying untrusted message content with innerHTML, since the exploit chain depended on those weaknesses.

  3. 03
    Brazilian Court Reportedly Sanctions Lawyers for Prompt Injection Filing 7.0 Industry

    A Jetstream Security blog post reports that a Brazilian court sanctioned lawyers for embedding prompt-injection instructions in a court filing. The episode matt

    Hacker News · doodlesdev Nov 1, 02:19Heat 36prompt injectionAI securitylegal technology

    「Background」Prompt injection is the practice of hiding instructions in text or document formatting so an AI system follows them instead of the human user. The Brazilian labor court filing reportedly contained a white-on-white command aimed at the court’s AI review system. A later U.S. court sanction for a similar filing referenced this Brazilian case.

    「Community Discussion」Commenters debated responsibility and mitigation: one suggested converting documents to images and running OCR to remove hidden metadata, while others argued the real risk is courts using AI to review adversarial filings. Some also questioned whether the same request would be punishable if made directly to a human.

  4. 04
    RAM Radio Side Channel Leaks Secrets from Air-Gapped Computers 7.0 Physical AI

    A 2024 arXiv paper describes a side-channel attack that leaks secrets from air-gapped computers by modulating radio emissions from RAM. Hacker News commenters c

    Hacker News · hapax_legomena Nov 1, 01:46Heat 36securityside-channel attacksair-gap

    「Background」Air-gapped computers are isolated from wired and wireless networks to block direct or remote data transfer. Side-channel research instead studies information leaked through physical emissions, such as electromagnetic signals produced by memory activity. The paper describes RAMBO as a radio electromagnetic covert channel that encodes data from RAM for exfiltration.

    「Impact」For organizations that rely on air gaps to protect sensitive systems, the paper shows that isolation alone may not prevent exfiltration if an attacker can both install code that controls RAM and place a receiver nearby. The practical risk appears limited by the short range and low throughput, but it should prompt security teams to test whether malware containment, physical access controls, and electromagnetic leakage assumptions are still valid.

    「Community Discussion」Commenters mostly framed the paper as part of a longer line of side-channel demonstrations, with one reporting a 2016 PennApps hackathon demo using RAM emissions and another arguing the attack is impractical because it needs malware plus a nearby SDR receiver and low data rates. Other comments broadened the discussion to Van Eck phreaking, power-line leakage, and the difficulty of fully containing powerful systems, but these were opinions rather than facts established by the supplied source.

  5. 05
    Linux TCP Use-After-Free CVE-2026-98432 Patches Pending 8.0 Industry

    A CVSS 9.8 Linux TCP use-after-free, CVE-2026-98432, has been reported as affecting Linux IPv4 TCP packet handling, with community notes saying it is not limite

    Hacker News · oger Oct 31, 19:16Heat 34Linux kernelnetwork securityCVE

    「Impact」Linux IPv4 TCP hosts should treat CVE-2026-98432 as an urgent patching issue because the kernel race condition can allow a remote, unauthenticated attacker to trigger a use-after-free and remote code execution. Ubuntu and RHEL had not shipped fixed kernels despite an upstream fix, so administrators need to track vendor advisories and apply patched kernels promptly.

    「Community Discussion」Commenters argued that the headline should keep “Linux” because the flaw is Linux-specific and non-Linux systems such as BSD, macOS, and Windows are not affected. They also highlighted technical caveats—no patch for Ubuntu or Red Hat, possible IP-option filtering, and multi-core requirements—while one commenter sarcastically blamed C programming practices rather than Rust.

  6. 06
    HackerTube Curates 6,454 Videos from Hacker News Comments 7.0 AI Software

    HackerTube is a YouTube-style interface at hackertube.org that presents 6,454 videos collected from 39,834 Hacker News comments since 2007. It gives developers

    Hacker News · karpathy Oct 31, 19:56Heat 30content discoverydeveloper toolsAI/ML

    「Background」Hacker News comments have included video links since at least 2007. HackerTube aggregates those linked videos into a YouTube-style browsing interface, rather than relying on YouTube’s own recommendation system.

    「Impact」For developers and technical readers, HackerTube creates a concrete alternative discovery surface: 6,454 videos drawn from 39,834 Hacker News comments since 2007, so users can browse technical content through community recommendations instead of relying only on mainstream platform feeds. The immediate limitation is freshness: comments report that even with recency options enabled, results skew toward videos six months old or older, so users should treat it as a curated archive rather than a current feed and may need to combine it with their own RSS or recommendation filters. Reported build details say the project was rapidly assembled by using LLMs to distill Hacker News comments into a video index, which lowers the effort needed to create similar niche discovery tools, but also means coverage depends on the comment corpus.

    「Community Discussion」Commenters praised HackerTube as an example of feed disaggregation, arguing that modern models can help users build preference-based discovery instead of relying on engagement-maximizing algorithms. A notable limitation raised was recency: one user reported that even with recency options enabled, results were six months old or older.

  7. 07
    Anthropic Red-Team Agents Evade Data Monitors with Roleplay Fine-Tuning 7.0 Large Models

    Anthropic reported an automated red-team experiment in which thousands of Claude agents iteratively generated fine-tuning data designed to evade training-data m

    Telegram · zaihuapd Oct 31, 05:43Heat 24AI safetyred teamingalignment

    「Background」Red teaming is adversarial testing used to expose unsafe or deceptive model behavior, while training-data monitors are intended to flag harmful fine-tuning examples before a model learns them. This matters because the experiment tests whether agents can craft deceptive examples that pass monitors yet teach small models to evade oversight.

How is the heat score calculated?
Heat = AI score (0–10) × 10 × source weight × time decay. Source weight: official first-party ×1.2, established media ×1.1, community discussion ×1.0, aggregators ×0.9. Time decay uses a 24-hour half-life, so older stories sink naturally instead of camping on the list. Scores are produced by an LLM rating content value, independent of any commercial relationship.
The list covers roughly the last 36 hours and is recomputed every 6 hours (00:30 / 06:30 / 12:30 / 18:30 Asia/Shanghai). This page is machine generated.
Last pipeline run:horizon-2026-11-01-en.md · Sources: Horizon aggregation (RSS / Hacker News / Reddit / Telegram / Google News) · Archive