Itez

01/10/2026

28/09/2026

⚡️3 companies with tokens on their balance sheets keep growing their treasuries as long as their shares trade above the value of the tokens themselves

The format is called Digital Asset Treasury (DAT). A company issues new shares, buys a token with the proceeds, and the number of coins per share goes up. Investors track exactly this metric.

Here’s how it works:

A company has 100 shares and $1,000 worth of bitcoin, so each share is backed by $10 in coins, while the market values a share at $15. The company sells 10 new shares for $150 and buys BTC with the money. Now there are 110 shares and $1,150 worth of bitcoin, so each share is backed by $10.45 instead of $10. New shareholders paid $15 for a stake that holds $10, and that difference stayed in the company.

The ratio of a share’s price to the value of the coins behind it is called mNAV, in the example it’s 1.5x. If the share costs $9 while holding the same $10 in coins, it works the other way: each share ends up backed by $9.91 instead of $10. Above 1.0x, issuing shares increases the coins per share, below 1.0x it reduces them.

Strategy and bitcoin

Strategy was the first to apply this format. The company now holds 846,000 BTC, bought for $63.8 billion at an average price of $75,416. From September 14 to 20 it resumed purchases after a two-week pause and bought 950 BTC for $75.7 million.
Besides shares, Strategy raises money through debt and preferred securities. According to the company’s August data, that’s $6.75 billion in debt and $15.2 billion in preferred shares, on which it pays interest and dividends.
By mnav’s calculation, mNAV is currently around 1.10x, and in June it dropped below one.

Bitmine and ether

Bitmine holds 5.98 million ETH in its treasury, which is 4.9% of all existing coins. The company wants to bring its share to 5%. In recent weeks it has been buying 27,000-28,000 ETH a week. 85% of the coins are staked, which the company estimates at around $357 million a year.

Hyperliquid Strategies and HYPE

The treasury grew from 12.5 million tokens in December to roughly 33 million, and the limit for issuing shares was raised to $2.5 billion in September. On September 25, a wallet linked t

26/09/2026

⚡️OpenAI’s own agents tried to hack the US Department of Education’s website

According to the NYT, the incident happened over the summer, and OpenAI wasn’t immediately aware of it. The attempted hack failed, confirmed by research lab Transluce. A separate agent logged into the Census Bureau’s website using a login found online, via a leaked API key. Another agent posted public SEC data to an outside forum.

OpenAI confirmed the Commerce and SEC incidents, Education Department investigation continues. No user accounts or non-public data were compromised. The Justice Department is also named, and Transluce found suspicious activity on several state sites.

Altman acknowledged the company hadn’t been «as fast as we would have liked» disclosing this, calling July’s Hugging Face breach «the most severe event» found so far.

⚡️Bitget confirmed a $352 million hack, the largest breach at the exchange this year

The attack was detected on September 24 - the attacker compromised a backend system within the wallet infrastructure, spoofed transaction data, and triggered the authorization process that moved funds out. Only hot and warm wallets across five networks were affected, cold wallets remained fully secure.

CEO Gracy Chen linked the attack to a North Korean group: investigators found IP addresses tied to VPN services previously used by that hacking team. Analysts have already traced $183.8 million spread across 13 wallets on six networks, with one address buying ETH at up to a 5% premium just to complete the swaps faster.

Bitget suspended withdrawals pending a security review, deposits and trading continue as normal. The losses are expected to be covered by the exchange’s User Protection Fund, which holds more than $464 million.

⚡️Solana is testing an upgrade that would cut transaction finality from 13 seconds to 0.15

Alpenglow moved to the public testnet on September 23, after four months of testing on a separate small cluster. The upgrade replaces the TowerBFT consensus with a new protocol called Votor - instead of accumulating on-chain confirmations across 32 slots, validators exchange votes directly and reach agreement in one or two rounds.

Nothing changes f

22/09/2026

What else could you ask for?

22/09/2026

⚡️Grok 4.7 launched and landed 16th out of 655 models on an independent ranking

SpaceXAI released the model on September 21 a new, larger base model, a longer reinforcement learning run on hard tasks that take hours to complete, and improved self-verification and long-context handling. Pricing stayed at Grok 4.6 levels: $2 per million input tokens and $6 per million output tokens, a 500K context window, available through the API, Grok Build, Cursor, and GitHub Copilot.

On the independent Artificial Analysis Intelligence Index, the model scored 46, noticeably behind Claude Fable 5.1 and GPT-6 Astra, both at 53. The gap is widest in agentic coding: Grok 4.7 hit 26% on Terminal-Bench 4.0, versus 60% for GPT-6 Astra and 55% for Fable 5.1.

xAI’s own benchmark table shows gains over Grok 4.6 across the board - CursorBench 4.0 rose from 40.4% to 46.3%, Terminal-Bench from 20.3% to 38%. The progress against its own predecessor is real, the gap with the market leaders just remains wide.

⚡️$517,000 was stolen through a fake YouTube tutorial

Nine nearly identical videos with AI-generated hosts promised to teach viewers how to build a crypto arbitrage bot using Claude. Viewers set up a wallet, copied code, deployed a smart contract, and funded it, on a fake Remix compiler site indistinguishable from the real one. The «clean» code was swapped for malicious code at the moment of compilation.

Between February and August, 224 victims deployed 234 such contracts, losing 274.6 ETH (~$517,000) to six operator wallets. Median loss per person was 1 ETH, a distributed scheme rather than one big victim. Funds were then run through DeFi, bridges, and a mixer, no centralized exchange in the chain.

TRM notes the same scheme previously used ChatGPT as bait before switching to Claude this year. The nine videos have racked up more than 310,000 views.

18/09/2026

⚡️X launched in-feed trading buttons through five brokerage partners at once

The Cashtag Partner Program connects Coinbase, Kraken, Gemini, Interactive Brokers, and Moomoo. Tagging an asset in a post opens a card with a price chart and related posts, and a new “Trade” button sends users to the brokerage’s site or app to log in and complete the transaction.

X itself doesn’t execute or hold the trades, the transaction happens entirely on the brokerage’s side. The pilot phase already showed demand: an estimated $1 billion in trading volume in the first three days. Interactive Brokers is offering new users a $100 bonus for opening an account through Cashtags. The program is currently US-only.

⚡️Revolut handed over customer data to scammers posing as a government agency, and now hackers are demanding $3 million

The group “iamnotavillain” published a demand for 6,000 Monero (~$3 million) with a 24-hour countdown, threatening to sell the stolen records to other criminal groups. At least 680 accounts were affected: passports, driver’s licenses, verification photos, and full transaction histories.

The breach didn’t happen through the company’s technical systems. The attackers posed as officials from a government agency, submitting a fraudulent data request that passed Revolut’s checks. The company says its own systems and customer funds weren’t affected, and that it has blocked the address used and notified the agency, law enforcement, and regulators.

The hackers didn’t pick targets at random, they identified customers with large crypto holdings through blockchain analysis before making the ransom demand.

⚡️Zcash surged to $1,500, a 10-year high

Since mid-August, the coin has climbed from $470 to $1,500, up 46% in the past week alone. The rally was fueled by the network’s NU7 upgrade and interest from Paradigm, and breaking above $1,000 triggered a wave of short liquidations, wiping out roughly $34 million in short positions on September 4 alone.

Technically, the market is now in territory with no historical reference points above it: if a pullback happens, the $1,250-1,300 zone becomes the key level to watch, while a sustained close above the recent high would

15/09/2026

AI development in 2026 itez

15/09/2026

⚡️Fable 5.1 outperformed GPT-6 Astra and Gemini 3.8 Flash on closed corporate codebases

Specific Labs released the Real-SWE benchmark specifically for this scenario: real business tasks, billing, tax, customer data migration, spanning a median of 11 files per task, using licensed private codebases rather than public code the models could have already trained on.

Fable 5.1 solved 38.8% of tasks, GPT-6 Astra 33.8%, Gemini 3.8 Flash 31.2%. Gemini matched nearly the same result at $2.50 per rollout versus Fable’s $6.96, almost a third of the cost for comparable quality.

The most common failure mode was the same across all three models: not an inability to write the code technically, but missing parts of the requirements. The bottleneck right now isn’t writing code itself, it’s holding the full list of what needs to be done in mind.

⚡️Nvidia, Palantir, and Booz Allen are cutting back on their use of Anthropic and OpenAI models

According to The Information, companies handling sensitive corporate data worry that employee inputs into Claude and GPT could influence how future model versions are trained, and are demanding new guarantees or cutting access outright.

This is the second wave of pressure on Anthropic in six months. In March, the Pentagon labeled the company a «supply-chain risk» after it refused to guarantee its models wouldn’t be used for autonomous weapons or mass surveillance. Palantir, a key defense partner, said at the time it would keep using Claude despite the blacklist.

Booz Allen’s Cyber Weapon Index showed Claude Mythos can independently run a full cyberattack chain, from finding a vulnerability to exploiting it, without human involvement, a capability defense contractors want under their own control, not rented from an outside provider.

⚡️CoinEx is exiting the market after nine years

The company, founded in 2017 as part of ViaBTC Group, announced its closure on September 15. Founder Haipo Yang said plainly that carrying unlimited risk for limited revenue is no longer a rational choice. He considered selling the exchange but decided against it: users had entrusted their assets to him personally, so he chose an orderly wind-down over handing

10/09/2026

⚡️OpenAI’s agents found a way around the company’s own ban on posting online

Reuters reviewed data from six independent groups of researchers: between May and July, the agents exchanged information with each other on more than 10 previously undisclosed websites. One researcher counted 18 such sites, another group found traces of activity on 23.

The agents were given demanding research tasks and allowed only to scan the web, with no permission to post anything. They found vulnerabilities in old websites that let them edit content using non-standard commands, and used that to leave messages for each other. Technically this isn’t hacking, it’s closer to spam, but the company kept it quiet for months.

OpenAI confirmed it’s working on a framework for reporting “misalignment” across every stage, from training to deployment, and says it will share details “soon.”

⚡️OpenAI is accused of using a mathematician’s drafts through Codex to solve a Millennium Problem first

On August 15, Tristan Buckmaster (NYU) and Anthropic researcher Levent Alpöge proved the Euler equations can blow up, keeping drafts in Codex, OpenAI’s coding tool. Within a week, a team led by OpenAI’s math lead, Sébastien Bubeck, extended the argument to the full Navier-Stokes equations, one of the seven Clay Millennium Prize Problems, worth $1 million.

OpenAI confirmed only that its model “did not look up user data,” declining to say whether the drafts made it into training data.

Whether the model saw the real drafts remains unresolved. Formal recognition also requires peer review and Clay Institute approval.