Russian Hackers Hunted, Chinese Chips Power Massive AI Model

Here's the part that should make you uncomfortable: the attackers don't need your password. They just need you to send them a number.

The US State Department is offering up to $10 million for information leading to the identification or location of a Russian state-linked hacking group that has been systematically compromising Signal and WhatsApp accounts since at least March. The targets aren't random — they're current and former US government officials, military personnel, political figures, and investigative journalists. Basically, exactly the people who chose encrypted messaging apps because they thought they were safe.

The operation works through social engineering, not brute force. Attackers send messages disguised as automated support communications from Signal itself, asking recipients to click a link or hand over a verification code. If the target complies, the attacker quietly links their own device to the victim's account — meaning every message sent from that point forward is visible to someone sitting in Moscow.

Signal's end-to-end encryption actually holds up here. Messages sent before the account was compromised remain protected. But the attackers evolved their tactics to close that gap too. A newer wave of phishing messages urges targets to back up their previous conversations by following a specific set of steps — steps that end with the target copying and sending over the long passcode used to encrypt those backups stored on Signal's servers. Hand that over and the past is no longer safe either.

The FBI has attributed the campaign to two Russian government-linked groups tracked as UNC5792 and UNC4221. The advisory was updated recently to reflect how significantly the operation has scaled and how the social engineering lures have grown more convincing. The fake messages are polished enough to mimic the tone and format of legitimate Signal notifications — complete with urgent language about account syncing failures and threats of permanent data loss.

This matters well beyond the individuals being targeted. Signal has become the de facto secure communications tool for journalists, activists, diplomats, and anyone operating in high-stakes environments. The fact that a state-level actor has found a way to compromise accounts at scale — without ever breaking the encryption itself — is a reminder that the weakest link is almost always the human at the keyboard.

The $10 million reward signals (pun intended) that the US government views this campaign as a serious ongoing threat rather than a one-off incident. For anyone using Signal professionally, the immediate takeaway is simple: if you receive any message asking you to back up your account, verify a code, or link a new device, treat it as hostile until proven otherwise. The real Signal team is not going to DM you.

A Chinese food delivery company just released one of the largest open-source coding models in the world, and it was built without a single Nvidia H100.

Meituan — best known in China for delivering your lunch in under 30 minutes — has open-sourced LongCat 2.0, a 1.6 trillion parameter agentic coding model that has reportedly been ranking at the top of OpenRouter's usage charts. That's a meaningful signal. OpenRouter aggregates traffic across dozens of AI models, so leading that leaderboard means developers are actively choosing it over well-funded alternatives from OpenAI, Anthropic, and Google.

The headline number is staggering. For context, GPT-4 is widely estimated to be somewhere in the 1 to 1.8 trillion parameter range, though OpenAI has never confirmed specifics. LongCat 2.0 sitting in that same neighborhood — and being released as open-source — is the kind of thing that would have seemed implausible from a food-tech company two years ago.

But the detail that will get the most attention in policy circles is where this model was trained. Meituan built LongCat 2.0 entirely on Chinese chips, sidestepping the advanced semiconductors that US export controls have been specifically designed to keep out of Chinese AI labs. The US has spent years tightening restrictions on Nvidia's most powerful GPUs precisely to slow down Chinese AI development. Meituan just published a 1.6 trillion parameter model and said, essentially, we didn't need them.

This doesn't mean export controls have failed entirely — training at this scale on domestic chips almost certainly took longer and cost more than it would have on cutting-edge Nvidia hardware. But it does suggest the gap is narrowing faster than some policymakers assumed, and that determined, well-resourced companies are finding ways to work around the bottleneck rather than waiting for it to be lifted.

The agentic focus of LongCat 2.0 is also worth noting. Coding assistants that can autonomously plan, write, debug, and iterate through complex software tasks are increasingly where the real competition in AI is playing out. It's a more commercially useful capability than raw benchmark performance, and it's what enterprise developers actually care about.

Meituan open-sourcing the model adds another dimension. The Chinese AI ecosystem has historically been more closed than its Western counterpart, but releasing a frontier-scale model publicly puts pressure on other labs — Chinese and American — to follow suit or explain why they won't. It also means the global developer community can now fine-tune, audit, and build on top of a model that was trained without Western chip infrastructure.

The story here isn't really about food delivery anymore. It's about what happens when export controls meet genuine engineering ambition — and which one blinks first.