Delhi Police warns against sharing photos on unknown AI tools amid blackmail risk

Delhi Police has issued a public advisory cautioning residents against uploading personal photographs to unverified AI-based tools. Authorities warn that such images can be misused to create fake profiles or facilitate fraud and blackmail. Users are urged to avoid clicking on suspicious or unverified links that could expose them to scams. Those who fall victim to cyber fraud are encouraged to report incidents through official police channels.

Colombo-Bound Flight Makes Emergency Landing in Chennai After Engine Failure

A flight bound for Colombo made an emergency landing at Chennai Airport following an engine failure. The aircraft was diverted to Chennai as a precautionary safety measure. The incident prompted an emergency response at the airport. Aviation authorities and the airline are expected to investigate the cause of the engine failure.

Data-Only Cyberattacks Pose Greater Risk Than Previously Understood

A publication by USENIX highlights that data-only attacks, which manipulate program data without altering code, are more accessible to attackers than widely assumed. Unlike traditional code-injection exploits, these attacks bypass many modern defenses by targeting runtime data structures. The research suggests that existing security mitigations may offer less protection than developers and organizations currently believe. The findings call attention to a growing gap between assumed and actual security postures in software systems.

FBI Probes Alleged 2TB Data Breach of Its Job Application Website

The FBI is investigating a cyberattack targeting its official job application portal, FBIjobs.gov. Hacker group ShinyHunters has claimed responsibility, alleging they stole over 2 terabytes of sensitive data belonging to FBI employees and applicants. The group says they exploited a vulnerability in Oracle's PeopleSoft system to gain access. Cybersecurity experts have raised concerns about the long-term risks to individuals whose data may be circulating among malicious actors.

Generative vs Agentic AI: How to Choose the Right Tool in 2026

As of September 2026, generative AI remains the preferred choice for content creation tasks such as drafting, summarisation, and code snippets, due to its lower cost and ease of review. Agentic AI, which is built on top of generative models, is better suited for multi-step, goal-directed workflows like bug fixing or data reconciliation, where results can be automatically verified. Despite growing interest, adoption of agentic systems remains limited — only 23% of organisations have scaled one, while 39% are still experimenting. Gartner warns that over 40% of agentic AI projects could be cancelled by end of 2027, citing rising costs, unclear business value, and insufficient risk controls. Experts recommend a default-generative approach, escalating to agentic only for well-defined, verifiable workflows where a machine can reliably grade its own output.

Pickup Crane Runs Over Teen Near College, Driver Flees Scene

A teenage girl named Prerna was fatally struck by a pickup crane while walking to her college. The incident was captured on camera and occurred just 200 metres from the college entrance. The crane driver was allegedly driving recklessly at the time of the accident. After running over the girl, the driver fled the scene without stopping to help. Authorities are investigating the incident based on the available footage.

Breeze TTS 2 Tops Open-Weight TTS Leaderboard but Bars Commercial Use

Breeze TTS 2, developed by BreezeBlue/RESONIA INC., reached the top spot among open-weight text-to-speech models on the Artificial Analysis leaderboard with an Elo score of 1215, surpassing ElevenLabs' Eleven v3 at 1175 as of September 11, 2026. Its hosted price of $34 per million characters is significantly cheaper than ElevenLabs' $100 for the same volume. However, while the model's inference code is licensed under Apache 2.0, the weights themselves are restricted to research and non-commercial use only, with commercial deployment requiring a separate written agreement from BreezeBlue. This makes Breeze TTS 2 a strong choice for researchers, hobbyists, and non-commercial builders with compatible GPU hardware, but unsuitable for businesses shipping paid products without prior licensing negotiations. ElevenLabs remains the recommended option for commercial use cases, given its unambiguous licensing, broader language support, and mature platform infrastructure.

DeepSeek V4.1-Flash Matches Top Closed Models on Coding at a Fraction of the Cost

DeepSeek released V4.1-Flash on 10 September 2026 with MIT-licensed weights, positioning it as a cost-efficient open-source model for coding agents. On the DeepSWE v1.1 benchmark, it scores 74.2, nearly matching Claude Opus 5 (74.0) and GPT-5.6 Sol (73.0), according to DeepSeek's own release figures. Its off-peak cached-input rate of $0.003 per million tokens makes it dramatically cheaper than rivals, with one worked example showing $0.15 versus $25 for Claude Opus 5 on a 100-request agent loop. However, the model significantly trails on hard reasoning tasks, scoring 36.8 on Humanity's Last Exam compared to Claude Opus 5's 56.3, making it less suitable for complex single-answer problems. From 14 September 2026, all V4-Pro API requests are automatically rerouted to V4.1-Flash and billed at the lower Flash rate.

Jev markets itself as a frontier AI model but is essentially a JSON classifier

Jev, a newly launched AI startup founded by ex-OpenAI researchers, claims to offer hallucination-free outputs and frontier-model capabilities. In practice, the product functions as a constrained JSON classifier that picks from predefined options, scores against criteria, or returns true/false judgments. Its 'no hallucinations' claim relies on constrained decoding — a standard technique already widely used by OpenAI, Anthropic, and open-source tools — which simply prevents the model from emitting tokens outside a fixed set. Within hours of Jev's launch, a developer independently replicated the core functionality using a small open-source model on Hugging Face. The product's real value lies in cheaper, faster classification and routing at scale, but that cost-efficiency pitch has been packaged as a more ambitious frontier-AI narrative.

Bengaluru's Kinetic-4B Beats Claude Haiku 4.5 on Tool Calling Accuracy and Speed

Conscious Engines, a Bengaluru-based AI lab, published benchmark results on 1 April 2026 showing its 4-billion-parameter Kinetic-4B model outperformed Anthropic's Claude Haiku 4.5 on a 300-sample tool-calling evaluation sourced from Composio. Kinetic-4B achieved 82.33% accuracy at 1.61 seconds p95 latency, compared to Haiku 4.5's 80% accuracy at 4.02 seconds and OpenAI's 120-billion-parameter GPT-OSS-120B's 76.33% at 7.99 seconds. Notably, Kinetic-4B's failed call rate of 4.67% was roughly half that of Haiku 4.5's 9.67%, which has practical implications for agent reliability and error-handling overhead. The model was fine-tuned from the open Qwen3-4B base using LoRA on a single rented GPU over approximately 4.5 hours, and the adapter has been released publicly on Hugging Face. Analysts caution that the benchmark is self-published and covers only a narrow task, and that general-purpose models like Haiku remain preferable for broader reasoning, coding, and summarisation workloads.

Pune Startup DTI Bets India Needs Court-Grade Proof Layer for Digital Signatures

Pune-based deep-tech startup Digital Trust Infrastructure India Limited (DTI), founded by insolvency professional Rupesh Kuche and incorporated in December 2025, is developing a device and platform aimed at making every digital signature independently verifiable in court years after it is executed. The company argues that India's existing digital public infrastructure — Aadhaar, UPI, eSign — authenticates identity at the moment of a transaction but does not preserve durable, court-grade evidence of what was actually authorized or consented to. This gap has real consequences: Indian banks and financial institutions reported ₹48,021 crore in fraud during FY2025-26, and a Supreme Court bench has described digital fraud as akin to 'dacoity'. DTI's proposed solution includes a hardware signing device and an evidence network called Ascert.in, both currently at the piloting stage with no publicly launched product yet. The startup contends that long-term transaction evidence preservation remains an unsolved problem at national scale, distinct from existing global content-provenance standards like C2PA.

India News Live Updates, 23 September 2026: TMC alleges attack on IT wing chief by BJP ‘goons’, slams assault on Humayun Kabir

CyberPanel, aaPanel, CloudPanel: Top Free ServerPilot Alternatives in 2026

ServerPilot charges between $5 and $20 per server plus up to $2 per app monthly, but lacks built-in email, DNS management, and a file manager. CyberPanel, built on open-source OpenLiteSpeed, offers email and DNS at no cost, making it a strong all-in-one free option. aaPanel provides a cPanel-style interface with a one-time Pro licence priced around $399, adding features like a WAF, sub-accounts, and WP Toolkit. CloudPanel is the closest free drop-in replacement for ServerPilot, supporting PHP, Node.js, and Python with a minimal memory footprint, though it excludes email and DNS. A typical setup of three servers and eight apps costs roughly $456 per year on ServerPilot Business, while CloudPanel handles the same workload at no charge.

IIT Bombay Professor Put on Leave After Student Dies by Suicide Post Exam Incident

A second-year student at IIT Bombay, Sahil Wakode, died by suicide after being caught allegedly using his mobile phone to access ChatGPT during an examination. The incident has sparked controversy, placing Professor Suryanarayana Doolla at the centre of the controversy. Following the student's death, the institute has sent the professor on leave pending further review. The case has raised serious questions about academic pressure and the handling of examination misconduct at premier institutions.

Pakistan's Fuel Import Reliance Puts Petrol Prices at Risk of Hitting PKR 1,000

Pakistan faces the prospect of petrol prices surging to PKR 1,000 per litre, driven largely by global fuel market volatility. The country's heavy dependence on imported refined fuel makes it particularly vulnerable to international price fluctuations. Any sharp rise in global crude or refined fuel prices is quickly transmitted to domestic pump prices in Pakistan. This structural dependency leaves Pakistani consumers with little buffer against external energy shocks.

US-Iran talks: What was discussed in Trump's 'good meeting' on the sidelines of UNGA

Nine pre-publish SEO checks that catch invisible errors before they cost you rankings

A structured SEO checklist published on DEV Community focuses on 'invisible' failures — issues that don't break a page visually but quietly damage search performance over time. The guide recommends running structural checks first, covering elements like H1 uniqueness, heading hierarchy, duplicate page titles, and canonical URL consistency with trailing slashes. It also flags that meta descriptions should be previewed in pixels rather than counted by characters, with critical content kept within the first 120 characters. Internal linking is highlighted as a fully controllable navigation signal, with every new page advised to link to at least two existing pages and receive at least one inbound link. Additional checks cover image dimensions and alt text for layout stability, plus accurate structured data markup to avoid misleading search engines.

90% of Top Homepages Have Open Graph Tags, But Many Are Broken or Incomplete

A developer audited Open Graph and Twitter Card meta tags across 78 well-known websites, finding 62 usable homepages to analyze. While 90% had at least one Open Graph tag, coverage dropped sharply for key fields — only 77% included og:image, and just 35% specified og:locale. Nearly one in five sites omitted twitter:card entirely, which prevents X (formerly Twitter) from rendering a visual card even when valid Open Graph tags are present. Two sites declared a large-image card type but supplied no actual image, with one site's og:image tag present in the HTML but completely lacking a content attribute — a subtle bug that passes most automated checks. Only 23% of sites with an image tag also included width and height dimensions, which can cause intermittent display failures when social platforms cannot pre-layout the card without downloading the image first.

Delhi's LSR College shifts classes online

Shipping via Hormuz stays below average