DD
MM
YYYY

PAGES

DD
MM
YYYY

spot_img

PAGES

Home Blog Page 65

Bitrefill Releases Post-Mortem after it Suffered Significant Cyberattack 

0

The cryptocurrency payments and gift card platform Bitrefill suffered a significant cyberattack. The company disclosed the incident in a detailed post-mortem. The attack began with a compromised employee laptop likely via malware or phishing, which allowed access to legacy credentials and parts of the internal infrastructure.

Attackers gained access to production keys, drained funds from hot wallets, and made unauthorized and suspicious purchases through supplier channels. Approximately 18,500 purchase records were accessed, exposing limited customer data such as: Email addresses. Crypto payment addresses; Metadata (e.g., IP addresses).

Some reports mention around 1,000 additional records with encrypted customer names potentially affected, but sensitive data like full payment details or passwords were not stored on Bitrefill’s systems; they use external providers for much of that. No widespread full account takeovers or major private key exposures for users were reported.

Attribution to North Korea’s Lazarus Group: Bitrefill and independent analyses pointed strongly to the Lazarus Group also associated with subgroups like Bluenoroff, a notorious North Korean state-sponsored hacking collective known for high-profile crypto thefts. Evidence cited includes: Similar malware patterns and tactics.

Reused infrastructure specific IP addresses, email addresses tied to prior attacks. On-chain tracing of stolen funds matching Lazarus and Bluenoroff behavior. The company collaborated with law enforcement and cybersecurity experts during the response. Bitrefill has since enhanced security measures, isolated affected systems, and resumed operations with added protections.

This incident highlights ongoing risks in the crypto space, especially from sophisticated state-linked actors targeting hot wallets and employee endpoints. No massive user fund losses were reported beyond the company’s hot wallets.

The Lazarus Group also known as Hidden Cobra, APT38, or subgroups like BlueNoroff and TraderTraitor is a North Korean state-sponsored cyber threat actor linked to the Reconnaissance General Bureau. Active since at least 2009, it blends espionage, destructive operations, and financially motivated theft—particularly targeting banks, cryptocurrency platforms, and exchanges to generate revenue and evade sanctions.

Their tactics, techniques, and procedures (TTPs) evolve but follow consistent patterns, mapped extensively in frameworks like MITRE ATT&CK. Here’s a breakdown of their core methods, with emphasis on cryptocurrency-related attacks (relevant to incidents like the recent Bitrefill breach). Lazarus heavily relies on human-targeted vectors rather than purely technical exploits.

Spear-phishing and social engineering — The most common method, often using fake job offers, investment scams, payroll themes, or collaboration lures. Victims download malware via attachments or links. Malware infects employee devices (laptops), exfiltrating credentials or keys.

In the Bitrefill case (March 2026), attackers started with a compromised employee laptop to steal legacy credentials, gaining access to production secrets and infrastructure.
Supply chain compromises — Trojanizing legitimate software, injecting malicious packages into open-source repositories (npm/PyPI), or exploiting upstream dependencies.

Watering hole attacks — Compromising sites frequented by targets. Use living-off-the-land techniques — Legitimate tools like PowerShell, WMI, or scheduled tasks for execution and persistence. Heavy obfuscation — Hex-encoding, variable mangling, software packing, and encrypted/encoded files to evade detection.

Multi-stage payloads — Initial droppers fetch further stages from C2 servers often via legitimate services like GitHub, Dropbox, or Slack for blending. Exploit vulnerabilities (zero-days or purchased exploits) in software.
Credential dumping. Registry modifications, run keys, or scheduled tasks for persistence.

System checks, time-based delays. Fileless techniques and masquerading as legitimate processes. Steal private keys, wallet seeds, or multisig approvals. Hot wallet drainage — Direct transfers from compromised wallets as in Bitrefill, where production keys enabled hot wallet drains and unauthorized purchases via suppliers.

In crypto hacks (Ronin, Harmony, Bybit, KuCoin, etc.): Focus on centralized exchanges, platforms via employee compromise or supply chain. Exfiltrate limited but valuable data (emails, addresses, IPs/metadata — similar to Bitrefill’s ~18,500 purchase records exposure).
Reuse infrastructure (IPs, emails, malware patterns) for attribution.

Lazarus shows high discipline: long reconnaissance, modular tools, and adaptation; shifting to open-source supply chains in 2025+. They fund North Korea’s regime, blending state goals with crime. Mitigation tips for crypto firms and users: Enforce MFA/hardware keys for all access.
Segment hot wallets, use cold storage.

Monitor for anomalous logins/credential use.
Train against phishing/social engineering.
Regularly rotate/audit credentials and patch systems. This group remains one of the most prolific threats in crypto, with billions stolen historically.

Tether Makes Breakthrough Advancing Local Private AI on Consumer Cell Phones

0

Tether; the company behind the USDT stablecoin has made a significant breakthrough in advancing local, private AI capabilities directly on consumer cell phones and other everyday devices.

Tether announced the launch of an enhanced version of their QVAC Fabric framework. This is described as the world’s first cross-platform LoRA (Low-Rank Adaptation) fine-tuning framework specifically optimized for Microsoft’s BitNet models (1-bit quantized large language models). The key innovation dramatically lowers memory and compute demands—achieving reductions of over 70% in some cases—allowing billion-parameter AI models to be fine-tuned (customized/trained on personal data) and run inference locally on hardware like: Modern smartphones (e.g., iPhone 16, Samsung Galaxy S25.

Consumer laptops and desktops. Standard GPUs including AMD, Intel, Apple Silicon, and mobile GPUs like Qualcomm Adreno or Apple Bionic. This enables fully on-device AI training and personalization without any cloud dependency, meaning your data never leaves your phone—maximizing privacy and enabling offline use.

Previous QVAC developments starting in late 2025 included tools like QVAC Workbench; a local AI app for running and training models and earlier Fabric versions for inference on heterogeneous hardware. This latest release builds on those by integrating BitNet’s ultra-efficient 1-bit architecture with LoRA, making high-level customization feasible on phones for the first time.

Tether’s engineers demonstrated real-world results, such as fine-tuning models up to 1 billion parameters in under two hours on flagship phones, and supporting up to 13 billion parameters in some cases. The framework is open-source, cross-platform, and positions Tether as a push toward decentralized, privacy-first AI infrastructure—countering centralized cloud providers.

This move aligns with Tether CEO Paolo Ardoino’s vision of “local private AI that can truly serve the people,” expanding the company beyond stablecoins into broader tech ecosystems, including potential integrations with mobile hardware partners.

It’s being hailed as a step toward truly personal, offline AI assistants that learn from your data securely in your pocket, with big implications for privacy, edge computing, and reducing reliance on Big Tech clouds. LoRA (Low-Rank Adaptation) is a very popular and efficient technique for fine-tuning large language models and other neural networks without needing to update every single parameter in the model.

It was introduced in a 2021 paper by Microsoft researchers (“LoRA: Low-Rank Adaptation of Large Language Models”) and has become one of the go-to methods for customizing big models like Llama, Mistral, GPT-style models, BitNet, and others — especially on limited hardware like consumer GPUs, laptops, or even phones as seen in recent frameworks like Tether’s QVAC Fabric.

Full fine-tuning of a large model is extremely expensive: A 7B parameter model has ~7 billion weights. A 70B model has ~70 billion. Updating all of them requires massive VRAM often 100+ GB even with tricks like quantization, huge compute, and long training times.

It also risks “catastrophic forgetting” where the model loses too much of its general knowledge. LoRA solves this by making fine-tuning parameter-efficient. When you fine-tune a large pre-trained model on a new task/dataset, the change in the weight matrices (let’s call it ?W) is often low-rank.

In other words, even though the original weight matrix W is huge and full-rank, the update needed for adaptation can be approximated very well by a much smaller, lower-dimensional change.

Instead of learning the full ?W which would be the same size as W, LoRA learns two tiny matrices A and B such that: ?W ? B × AWhere:Original weight matrix in a layer: W (size d × k, e.g., 4096 × 4096 in many transformers). A is initialized randomly (usually with small values), size d × r. B starts as zeros (so ?W starts at zero, no change at the beginning), size r × k.

r is the rank — a small number you choose very important hyperparameter, typically 4, 8, 16, 32, or 64 — much smaller than d or k. During forward pass, instead of just using W, the model computes: W’ = W + (B × A) or more precisely: h = Wx + (B × (A × x)) scaled by some factor ? The original W stays frozen (never updated, no gradients).

Only A and B are trained ? number of trainable parameters drops dramatically (often 0.1%–1% of full fine-tuning). Quick math example Suppose a weight matrix W is 4096 × 4096 = ~16.8 million parameters. With LoRA rank r = 16:A: 4096 × 16 = ~65k params. B: 16 × 4096 = ~65k params. Total trainable: ~130k (instead of 16.8M) ? ~0.8% of original.

Yet in practice, LoRA with reasonable rank often matches or even beats full fine-tuning quality on many tasks. Key advantages of LoRAMuch lower memory — you can fine-tune 70B models on a single 24GB GPU or even larger with quantization like QLoRA. Faster training — fewer parameters to update.

Small adapter files — a LoRA for a 70B model is often just 10–200 MB instead of 140 GB. Easy to merge/switch — you can keep many LoRAs (one per task/personality/style) and merge them into the base model or swap them at inference time with almost no overhead.

No extra inference latency after merging though some implementations keep a tiny overhead if not merged. Works great with quantization. Common hyperparameters in LoRArank (r): The bottleneck size. Higher = more expressive (but more params and memory). Start with 8–32. alpha (?): Scaling factor for the update (often ? = 2×r or similar). Controls how strong the adaptation is.

Sometimes added to A/B matrices. target modules: Which layers to apply LoRA to usually attention Q, V, sometimes O, MLP, etc. In frameworks like Hugging Face PEFT, bitsandbytes, or Tether’s QVAC Fabric optimized for BitNet and mobile, you just set these and it handles injecting the adapters.

In short: LoRA lets you “personalize” massive AI models very cheaply and privately — exactly why it’s a breakthrough for running customized, local AI on phones and consumer devices without sending your data to the cloud.

PayPal Expands Access to its Dollar-backed Stablecoin to 70 Markets Worldwide

0

PayPal has recently expanded access to its dollar-backed stablecoin, PayPal USD (PYUSD), making it available in 70 markets worldwide.

This marks a significant global push for PYUSD, which was previously limited primarily to users in the United States and United Kingdom since its launch in 2023. The expansion adds around 68 new countries and regions, covering parts of Asia-Pacific, Europe, Latin America, North America, South America, Africa, and other areas.

Users in these 70 markets can now buy, hold, send, receive, and in many cases earn rewards on PYUSD directly within their PayPal accounts. PYUSD enables faster settlements, lower-cost cross-border transfers (compared to traditional methods), and easier participation in the global economy by reducing fees and currency conversion issues.

The stablecoin is federally regulated in the US, fully backed by USD deposits and Treasuries, and supports transfers to third-party wallets or conversion to local currencies in supported areas. PayPal operates in roughly 200 countries overall, so this rollout covers a subset, with additional markets expected to gain access in the coming weeks.

Examples of Included Countries and Region: Colombia, Peru, Guatemala, Honduras, Panama, Costa Rica, Dominican Republic (Latin America/South America). Uganda (Africa). Singapore (Asia-Pacific). United Kingdom and United States (pre-existing). Others like Faroe Islands and Greenland.

PayPal’s head of crypto, May Zabaneh, emphasized that this move provides “faster access to funds, lower-cost ways to send money across borders, and a more direct path to participating in the global economy.” PYUSD’s market cap has grown substantially reflecting increasing adoption. Users can check availability directly in the PayPal app as rollout may vary slightly by location.

This positions PayPal as a major player in bridging traditional finance with stablecoins for everyday global payments. This move transforms PYUSD from a primarily U.S./U.K.-focused product (launched in 2023) into a more globally accessible, dollar-backed stablecoin integrated directly into PayPal’s ecosystem of hundreds of millions of users.

Traditional international remittances often involve high fees (5-7% in many corridors) and multi-day settlement times. PYUSD enables near-instant or minutes-fast transfers with significantly lower costs, particularly benefiting regions like Latin America. This directly addresses pain points in high-remittance economies.

In many emerging markets, users gain easier access to USD-pegged value storage without immediate currency conversion. Eligible holders can earn rewards similar to the ~4% in the U.S., introducing a “balance + earnings” model in PayPal wallets—essentially turning stablecoin holdings into a yield-bearing option.

Individuals in these 70 markets can buy, hold, send, receive, and transfer PYUSD to external wallets, reducing friction for freelancers, small businesses, or families relying on international payments. Merchants accepting PYUSD get proceeds available in minutes vs. days/weeks traditionally, aiding cross-border operations and global commerce participation.

Reduced reliance on costly legacy systems could boost margins for e-commerce sellers and international suppliers. As a federally regulated (U.S.) stablecoin fully backed by USD deposits and Treasuries, PYUSD offers compliance-friendly entry into crypto for businesses wary of unregulated options.

With PYUSD’s market cap already around $4-4.1 billion, this rollout aims to build the “liquidity, utility, and ubiquity” needed for mainstream use. It positions PayPal as a key player in competing with dominant issuers like Tether (USDT) and Circle (USDC) by leveraging its massive user base ~430 million active accounts.

PayPal’s push validates stablecoins for everyday global payments, accelerating integration with traditional finance. It highlights how fintech giants are racing to capture remittance and settlement flows—especially as competitors like Visa/Mastercard build blockchain layers.

While promising, success depends on regulatory alignment in each market competition from established stablecoins, and user trust in centralized custody. Rollout is phased, with more markets expected soon, but full global coverage (PayPal operates in ~200 countries) remains incomplete.

This expansion accelerates the convergence of fiat payments and crypto rails, potentially reshaping cross-border finance by making dollar-based digital money more inclusive, efficient, and rewarding—especially in underserved regions.

PayPal’s head of crypto, May Zabaneh, framed it as driving “commerce forward for everyone” by tackling outdated systems. Adoption will hinge on real-world usage growth in the coming months.

Joe Kent Resignation from NCTC Projects Falling Inner Support within Trump’s Administration 

0

Joe Kent, the Director of the National Counterterrorism Center (NCTC), has resigned in protest over the ongoing U.S. war with Iran under President Donald Trump. Kent, a decorated veteran, former CIA paramilitary officer, and longtime Trump supporter, announced his resignation on March 17, 2026.

In a public letter posted on X, he stated that he could not “in good conscience” continue supporting the conflict. He claimed:Iran posed no imminent threat to the United States. The U.S. “started this war due to pressure from Israel and its powerful American lobby.” The administration had been influenced by a “misinformation campaign.”

He urged Trump to “reverse course” and expressed concern about sending American troops into an unnecessary war. This marks the first high-profile resignation from within Trump’s administration over the Iran conflict, which appears to involve U.S.-backed or -led military actions including strikes alongside Israel, now in its third week or so based on reports.

Kent’s departure highlights emerging divisions within Trump’s base and national security circles regarding the justification and escalation of the war.

AP News: Emphasizing his view that Iran was not an imminent threat. Noting the protest against the U.S. war on Iran. Trump has reportedly downplayed the resignation, with some coverage indicating he appears unfazed, though figures like Steve Bannon have commented on potential ongoing fallout.

The resignation coincides with other developments in the conflict, such as Israeli strikes killing senior Iranian figures. It has sparked discussions about internal dissent, potential risks to U.S. security focus, and broader implications for the administration’s foreign policy.

Steve Bannon has commented on the fallout from Joe Kent’s resignation as Director of the National Counterterrorism Center (NCTC), framing it as a serious issue that “isn’t going away” and signaling deeper divisions within the MAGA/Trump base over the Iran war.Key points from reports and coverage: Bannon warned that the resignation and Kent’s claims represent ongoing tension.

He emphasized the need for “answers” as the fallout continues to unfold, suggesting this highlights unresolved questions about how the conflict escalated and who influenced the decision-making. This aligns with Bannon’s broader “America First” skepticism toward Middle East interventions, particularly any perceived shifts away from Trump’s original objectives.

Earlier War Room commentary from Bannon (pre-resignation) criticized how the conflict “shifted” due to Israeli actions, arguing Israel is a “protectorate” whose moves must align with U.S. goals—not dictate them. Some coverage notes indirect support or sympathy from Bannon-aligned figures.

For instance, Grace Chong; a close Bannon associate and War Room contributor praised Kent’s stand as “REAL COURAGE” and true “putting America first.” However, Bannon himself has not issued a direct, lengthy public endorsement or condemnation of Kent personally in the immediate aftermath. His comments focus more on the broader implications for the administration, the base, and the war’s direction—warning of persistent dissent rather than dismissing or attacking Kent outright.

Trump allies and administration figures have largely turned against Kent; calling him a “crazed egomaniac” or similar in smears, while isolationist voices see his exit as exposing internal rifts. Bannon appears to position the episode as evidence that criticism from the anti-interventionist wing won’t fade quietly.

This comes amid mixed reactions: Trump downplayed the resignation, some MAGA hardliners called it “good riddance,” and others including Bannon’s circle view it as a principled stand against unnecessary escalation. Bannon’s take underscores potential ongoing “fallout” in terms of base fractures, demands for transparency, and questions about foreign influence on U.S. policy.

No major new direct Bannon statement has surfaced in the last day, but his War Room platform has historically amplified similar concerns about the Iran conflict’s trajectory.

X Has Begun Rolling Out a “Dislike” Button Specifically for Replies 

0

X has begun rolling out a “dislike” button specifically for replies, along with related features to help filter and manage reply quality. It’s currently in a limited, phased rollout; server-side, so not everyone sees it yet—update your app and check replies under posts.

The button appears as a broken heart or thumbs-down icon next to the like (heart) button on replies. Importantly, dislike counts are private—they aren’t shown publicly to anyone. Instead, the signals feed into X’s internal ranking algorithm to: Push low-quality, spammy, irrelevant, AI-generated, or troll replies lower in threads.

Promote more relevant and high-quality responses to the top. This helps clean up conversations without creating public “dislike piles” that could encourage negativity or brigading similar to concerns with YouTube dislikes in the past.

The feature reportedly went live quickly after X’s Head of Product, Nikita Bier, responded to a user suggestion with “Give me 60 seconds”—and screenshots started appearing within minutes, showing how fast X can ship tests these days. Alongside this, X is introducing or testing region-based reply restrictions, where users can limit who can reply to their posts based on country or region.

This is positioned as an anti-spam tool: It aims to reduce profitability of spam and scams by making it harder for bots or bad actors in certain areas to flood threads. Combined with the private dislike signals, it’s part of a broader push to curb phishing surges, ragebait, and low-effort replies.

Some reports mention additional feedback prompts when disliking (e.g., categories like “Spam,” “AI generated,” or “Misleading”). This rollout coincides with ongoing efforts to improve reply quality amid complaints about spam and scams on the platform. Earlier code leaks from mid-2024 onward had hinted at downvote/dislike mechanics, but this marks the actual user-facing deployment starting today.

Reactions on X are mixed—some users are excited about better threads, while others worry about potential abuse like coordinated dislikes. It’s still early and in testing, so expect tweaks based on feedback. If you’re seeing it or not, feel free to share what your experience has been.

The rollout of X’s “dislike” button is still in its very early, limited testing phase. As a result, observable impacts are mostly anecdotal from initial user reports and speculation, rather than large-scale data or long-term studies. Here’s a breakdown of the potential and emerging effects based on how the feature is designed and early reactions.

The core goal is to use private dislike signals plus optional feedback categories like “Spam,” “AI generated,” “Misleading,” or “Irrelevant” to demote low-effort, bot-generated, spammy, or off-topic replies in the algorithm. Early users who have access report that this could make conversations feel more relevant and less noisy—similar to how Reddit’s downvotes help surface better content without public shaming.

Many describe it as a “silent killer” for trash replies, potentially leading to higher-quality discussions over time. Since dislike counts are not public unlike YouTube’s old public dislikes or Reddit’s visible scores, there’s less motivation for coordinated dislike brigading or posts designed purely to provoke backlash. This design choice aims to avoid turning threads into negativity contests while still giving the algorithm useful signals to prioritize better replies.

Better Spam and Scam Control

Combined with other tools like region-based reply limits, it could make it harder for low-quality actors (bots, phishing attempts, AI slop) to dominate threads, indirectly improving the overall user experience and trust in conversations. Coordinated groups could still misuse the feature to bury dissenting or unpopular (but valid) replies, especially in polarized topics.

Without transparency on how signals are weighted, some worry it might amplify majority opinions or algorithmic quirks over time. Critics argue it could subtly chill certain types of replies; if the algorithm starts heavily demoting them. A few early reactions compare it to turning X “into Reddit,” where downvoting sometimes suppresses minority views or creates echo chambers.

No public counts mean users can’t see if a reply is widely disliked, which some see as a pro (avoids pile-ons) but others view as a con (harder to gauge community sentiment directly). It’s server-side and phased, so not everyone sees it yet. Some users are excited and already “smashing” the button on spam, while others haven’t encountered it and are skeptical or neutral.

This seems like a thoughtful evolution toward better conversation hygiene without repeating past mistakes. If it scales successfully, it could meaningfully reduce spam/AI slop and elevate thoughtful replies, making X threads more enjoyable—especially in high-engagement posts. However, it’s too soon for definitive impacts; expect tweaks based on data from this test.