Entities

Entities collected from this article, grouped for quick reference and copying. Inclusion does not establish malicious activity. Article links appear separately at the bottom as references.

The security industry stopped relying primarily on signature-based detection years ago. Static indicators like file hashes and specific binaries were always trivial for a threat actor to change. Detection moved to behavior, the how of an attack rather than the what, because behavior is more expensive to fake. The shift from indicators of compromise to indicators of attack is the ground modern defense stands on.

Russia’s COZY BEAR just used AI to go after what is left of the old model.

Anthropic disclosed a Russian state operation it tracks as GTG-20006, with tradecraft and targeting consistent with COZY BEAR, part of the Russian SVR.

COZY BEAR built AI agents to watch its own malware in the wild. The second a security product detects the malware, the agents rebuild it on the spot. Automatically. Over and over, modifying and recompiling until nothing flags it anymore, then redeploying it. The human barely touches a thing.

Static-indicator detection rests on one assumption. When a defender catches a piece of malware, the threat actor has to do real work to make a new variant. The small amount of work, and the window it offered, was the defender’s leverage. Detection imposed a cost, and the cost slowed the threat actor down.

COZY BEAR has now automated the rebuild, so the cost is mere seconds of compute. A defender detects the static artifact, and before a signature can be engineered, the AI has already shipped a version the indicator will not catch. Detection at the artifact layer is no longer the end of the fight. It is the trigger for the next variant.

This hits the static-indicator layer hard, but the tradecraft underneath still has to execute. The phishing, the DNS hijacking, and the ClickFix lures still look like something, which is exactly why behavioral detection matters more now, not less.

The same operation used AI to research and register phishing domains, stand up the hosting, send the emails, and watch the command-and-control channels for hits. It hijacked WhatsApp accounts by linking them as companion devices and quietly bulk-exported Ukrainian and Russian conversations while hiding the read receipts.

It even went after surveillance cameras by finding authorization flaws in camera-streaming services, then harvested tokens to watch the live feeds of victims. One operation stole an entire national identity database of 300,000 records, plus the commercial registry for half a million companies. All of it targeted Ukrainian and European governments, defense, and diplomacy.

AI has collapsed the gap once separating a well-funded state operation from a lone threat actor, and defenders should care about this most. The techniques are not getting magically more exotic. Intense labor is merely disappearing. Reconnaissance, exploit development, infrastructure, and data processing, all the expensive human work surrounding the attack, are being compressed into compute.

What COZY BEAR did here is not all that novel. It is a simple, modern, effective workflow, and a workflow can be copied. The self-rebuilding malware a Russian intelligence service just ran is the kind of capability trickling down to the eCrime ecosystem within days, weeks, or at worst, months, just as every other capability has moved from nation state to eCrime.

Anthropic caught this and disrupted it. It published the details, which is the argument for the labs actually reporting abuse of their own tools rather than burying it.

The same class of AI people use for ordinary work can be pointed at a defender and told to keep changing until the defender cannot see it, and it will do exactly that. COZY BEAR keeps showing how quickly it adapts.

Yesterday, detection bought defenders time. Today, detection may become a feedback signal, teaching an automated threat actor what to change next to sharpen its evasion.