Duane 'Keffe D' Davis convicted of murdering Tupac Shakur after 28-year wait

Duane 'Keffe D' Davis has been found guilty of first-degree murder for orchestrating the 1996 killing of rap icon Tupac Shakur. The verdict came nearly three decades after Shakur was shot dead in Las Vegas. Prosecutors built their case around multiple statements in which Davis himself admitted involvement in planning the fatal attack. His defense team contended that those admissions were fabricated, though no witness supported that argument. Davis now faces the possibility of spending the rest of his life in prison.

Developers Cancel AI Coding Subscriptions Over Unpredictable Usage Limits

Developers on Hacker News have been publicly abandoning paid AI coding tools like Claude Pro and Codex, citing usage limits that run out faster than expected and without clear explanation. Multiple users reported exhausting their monthly allowances within minutes or hours of starting complex tasks, prompting mid-project switches to competing services. The core frustration is not the tightness of the limits but their illegibility — subscribers cannot predict how limits are measured or when they will run out, making planning impossible. Some users concluded that paying directly through usage-based APIs is more reliable than flat subscriptions, since per-token billing at least makes costs transparent and predictable. The pattern documented across the week's discussions points to a recurring cycle: a heavy user hits an invisible wall mid-task, and that friction — rather than model quality — is what drives the cancellation.

Spring Ring Vishing Campaign Exploits Microsoft Teams to Target Corporate Networks

A threat campaign dubbed Spring Ring uses fraudulent Microsoft 365 external tenants to impersonate corporate IT staff via Microsoft Teams chat and voice calls, according to a Unit 42 report published August 31, 2026. Attackers trick employees into granting remote access through RMM tools like Quick Assist or running malicious executables delivered via customized S3 URLs. In more advanced intrusions, the campaign deploys an obfuscated PowerShell RAT and attempts NTLM Relay attacks against domain controllers using the PetitPotam technique to escalate privileges. The attack chain can result in persistent access, hidden browser instances with sideloaded extensions, and lateral movement across internal networks, though NTLM Relay was blocked in observed cases. Security teams are advised to restrict external Teams communication, enforce RMM tool approvals, enable PowerShell and AMSI controls, and harden NTLM relay defenses including SMB signing.

ServiceNow Patches Four Critical Flaws Including Three CVSS 10.0 Vulnerabilities

ServiceNow has released patches for four severe security vulnerabilities affecting its AI Platform and Now Platform, three of which carry the maximum CVSS score of 10.0. The flaws include unauthenticated code injection via the GraphQL Composite Data API (CVE-2026-18885), a privilege escalation bug in configuration image upload processing (CVE-2026-18886), SQL injection in dynamic ORDER BY clauses (CVE-2026-74820), and a sandbox escape vulnerability (CVE-2026-6876). All four can be exploited remotely without authentication or user interaction, potentially allowing attackers to execute arbitrary code, manipulate databases, escalate privileges, and compromise integration credentials. No active exploitation has been confirmed in the wild as of the disclosure date. Organizations are advised to apply the available hotfixes immediately and restrict external access to ServiceNow administration, API, and upload endpoints.

TerminalFix Malware Uses Fake CAPTCHA to Tunnel Into Corporate Networks via WebSocket

Microsoft has warned of a high-severity attack campaign called TerminalFix, disclosed on August 31, 2026, in which threat actors compromise websites to display convincing fake Cloudflare CAPTCHA pages that trick users into pasting malicious PowerShell commands into their terminals. Once executed, the malware extracts a ZIP archive, performs DLL side-loading through the legitimate LockScreenContentServer.exe, and reconstructs payloads hidden inside PNG image files using steganography. The infection establishes dual persistence via a Windows Run registry key and a 60-minute scheduled task, while also enumerating Active Directory and internal servers for reconnaissance. A Python-based reverse tunnel client then relays arbitrary TCP traffic over an encrypted WebSocket connection to an external command-and-control server, effectively turning the victim's endpoint into a pivot point into the internal network. Defenders are advised to enforce PowerShell Constrained Language Mode, block execution of known binaries from non-standard paths, and monitor for suspicious WebSocket connections to domains such as gitnow.dev.

Chinese Hackers Hijack Cisco Routers and TACACS Servers for Long-Term Espionage

A Chinese threat group known as Fire Ant has compromised Cisco IOS XR routers, TACACS+ authentication servers, and Linux management hosts to build persistent espionage infrastructure. The attackers deployed custom malware including BridgeAgent and the TacTap tool to steal TACACS and SSH credentials, enabling them to impersonate legitimate network administrators. They used hidden GRE tunnels, suppressed system logs, and captured network traffic into PCAP files that were exfiltrated to external FTP servers. The campaign targeted management plane infrastructure to gain a foothold for lateral movement toward high-value networks, including critical infrastructure. Security firm Sygnia recommends isolating management planes, cross-validating router states using independent sources, and rebuilding compromised systems from verified clean images.

Alteon, Indian startup building year-long autonomous aircraft, wins Lachy Groom backing

Indian startup Alteon is developing autonomous aircraft designed to remain airborne for up to a year by harvesting wind energy. The company was founded by a 20-year-old entrepreneur. Prominent investor Lachy Groom has backed the venture, signaling confidence in its technology. The aircraft are intended to operate autonomously, potentially offering a novel approach to long-duration aerial missions.

Explainer | Meta’s $17.1 billion toll shifts focus

Oracle trims 3,000 roles; Microsoft puts 500 on PIP

WestBridge books 2.6x return on PhysicsWallah exit

Times Square stabbing: 1 killed, 1 critically injured; suspect shot dead by NYPD

Indian cos chase ₹25,000 cr of D-St money in Sep

Why Mercury Auto-Rejects RU/BY Passport Holders and What Founders Can Do

Mercury Bank sends automated rejections to founders holding Russian or Belarusian passports due to an asymmetric risk calculation: OFAC violation penalties can reach $1 million per transaction, while revenue from a single such account is at most a few hundred dollars annually. The decisioning system flags RU and BY passports at the KYC stage regardless of the applicant's actual residence, business structure, or legal standing. Appeals rarely succeed because human reviewers operate within the same policy framework and are not re-evaluating eligibility — they are confirming the original flag was applied correctly. Founders with a second citizenship from countries like Israel, Armenia, or Georgia are advised to use that document as their primary KYC identity instead. Those without an alternative passport can explore US banking options such as Relay Financial, Novo, or Bluevine, which accept a broader range of non-US founders including some RU and BY passport holders with non-sanctioned residency.

WP-CLI's db check and db optimize commands can prevent silent WordPress database failures

WordPress databases accumulate bloat over time through spam comments, expired transients, post revisions, and orphaned plugin tables, which can cause corruption and performance issues. WP-CLI's wp db check command runs a MySQL health scan on every database table, helping catch corruption before it causes visible site errors. The wp db optimize command defragments tables and reclaims disk space from deleted rows, with behavior varying depending on whether the storage engine is InnoDB or the older MyISAM. Both commands are best run during low-traffic periods, and are especially useful after bulk deletions, before major updates, or as part of routine maintenance. If wp db check reveals corruption, wp db repair is available as a recovery step, though a fresh backup should always be taken before attempting any repair.

How to Build a Smarter Image Isolation Pipeline for SaaS Product Catalogs

Product catalog teams face a recurring trade-off between automated background removal and manual cropping when processing seller-uploaded images. Automated removal works well for high-volume, clean-background photos, while manual review is better suited for complex edges, transparent materials, or high-value hero images. A confidence-gated queue — rather than a fixed preference for one tool — helps route each image to the right processing path. Storing the original file as immutable and attaching metadata such as crop coordinates, confidence scores, and policy version to each result makes quality disputes easier to resolve. A simple decision table shared across engineering, catalog operations, and support teams provides a common vocabulary for handling exceptions in the review queue.

How strict audio validation silently killed emotion in a TTS training pipeline

A developer building a training dataset for an emotion-expressive text-to-speech model discovered that their automated quality control process was systematically filtering out the most emotionally rich audio clips. The pipeline used OpenAI's Whisper to validate generated speech against scripts, but emotion-laden audio — featuring trembling, pitch variation, or fading — was harder for Whisper to transcribe accurately and failed validation more often. As a result, flat, monotone takes consistently passed the filter, biasing the training corpus toward emotionless speech. Cosine similarity measurements confirmed the problem: bulk-generated clips scored 0.77–0.94 against a neutral voice, while a well-crafted emotional sample scored just 0.164. The fix involved a two-stage selection process — first filtering for transcription accuracy, then ranking surviving clips by emotional distance from neutral to pick the most expressive take.

Study Finds 38% of eBay Sold-Listing Rows Filtered Out Before Price Analysis

A 100-product study run through the CompSniper sold-price API found that relevance cleaning removed 37.87% of raw eBay sold-listing rows, reducing 19,220 parsed rows to 11,942 usable ones. The study covered five categories — smartphones, gaming, cameras, audio, and collectibles — each with 20 products, all queried under identical settings on eBay.com. Of the 100 products tested, 34 saw their median sale price shift by at least 10% after cleaning, with changes moving in both directions: 73 medians rose and 21 fell. The largest swing was a 283% increase for the GoPro Hero 12 Black, while the Google Pixel 7 Pro saw the steepest drop at nearly 53%. Collectibles and luxury items had the highest removal rate at 47%, though the median absolute price change across all categories remained modest at 4.2%, suggesting noise listings often cluster near the product's actual price range.

Apple Claims Former Employee Destroyed Evidence Tied to OpenAI Data Theft

Apple has presented what it describes as compelling evidence against a former employee accused of stealing confidential company data. The tech giant alleges the individual was connected to sharing proprietary information with OpenAI. According to Apple, the ex-employee took steps to destroy evidence after becoming aware that an investigation into the matter had been launched. The case highlights growing concerns around data security and the movement of personnel between major AI-focused companies.

Three Vision AI Models Tested on Product Catalogs: Relationship Mapping Remains Unsolved

A hands-on benchmark evaluated Mistral OCR, DeepSeek V4 Flash Vision Exp, and Qwen3-VL-32B-Instruct on a 138-page image-only product catalog to test their ability to convert visual data into structured, queryable records. While all three models accurately identified campaign prices, none successfully resolved shared price blocks or correctly extracted rotated SKU codes from a key test page. DeepSeek hallucinated incorrect product codes, while Mistral misclassified a list price as a campaign price, highlighting that character recognition is not the core challenge. A notable finding was that Qwen3-VL's extraction was actually successful, but a JSON field naming mismatch in the automated evaluator falsely reported a 0% accuracy score. The team concluded that evaluation pipeline integrity and human-in-the-loop workflows are as critical as model performance when building production-grade catalog extraction systems.

How a Ryan Gosling Tweet Forced a Flask App to Handle 10,000 Requests Per Second

A developer's small analytics backend, built with Flask on a 512MB Heroku dyno, collapsed under a sudden surge of 50,000 concurrent users after Ryan Gosling tweeted a link to their side project. The root cause was Flask's default thread-per-request model, which caused unbounded memory growth and repeated out-of-memory crashes on the limited server. The developer rebuilt the system using asynchronous I/O, a bounded SQLite connection pool, and WAL (Write-Ahead Logging) mode to safely manage concurrency without adding hardware. Key architectural constraints included capping connections at 200, setting fast failure timeouts, and offloading non-critical work to the client browser. The redesigned system ultimately sustained 10,000 requests per second on just 8GB of RAM, serving as a practical case study in resource-constrained backend engineering.