THE MALWARE THAT THINKS FOR ITSELF
.png)
What happens when malware stops waiting for the hacker?
For years, malware followed orders. A hacker compromised a computer. The hacker sent commands. The malware executed them.
But what if the malware could ask AI what to do next?
Security researchers have now uncovered a Windows malware called CLOSEDQUORUM that takes that idea a step further.
Instead of relying entirely on a traditional attacker-controlled command-and-control server, CLOSEDQUORUM can consult up to four AI models and use their responses to decide what action to take next.
The models involved include Google Gemini, DeepSeek, Qwen and Mistral.
The malware essentially creates an AI voting panel.
And that changes the conversation. (blog.talosintelligence.com)
THIS IS NOT JUST “AI-ASSISTED” MALWARE
We've already seen cybercriminals use AI to:
Write phishing emails
Generate malicious code
Translate scam messages
Automate reconnaissance
Create convincing social-engineering content
But CLOSEDQUORUM introduces something different.
The AI models are being used as part of the malware's decision-making process.
After compromising a Windows machine, the malware gathers information about the host, including details such as the Windows version and administrator status.
It then sends that context to configured AI services.
The models are asked to select what should happen next.
The available decisions include:
STEAL
INJECT
PERSIST
MOVE
The malware then counts the responses and executes the action receiving the most votes.
FOUR AI MODELS. ONE DECISION.
Think of it like a criminal committee.
One AI says:
“Steal credentials.”
Another says:
“Establish persistence.”
Another recommends:
“Inject code.”
The malware collects the responses and selects the winning action.
If there is a tie, the implementation has a defined priority order, with DeepSeek receiving the tie-breaking preference.
This means the malware isn't simply asking an AI to write code.
It is using AI as part of the operational decision loop.
Cisco Talos described CLOSEDQUORUM as the first publicly documented Windows implant it had identified that delegates tactical command-and-control decisions to a panel of AI models. (blog.talosintelligence.com)
WHAT CAN IT ACTUALLY DO?
The capabilities observed in the analyzed code are serious.
1. STEAL
The malware can attempt to collect:
Windows credentials from LSASS memory
Saved browser credentials
Chrome, Edge and Firefox data
Cryptocurrency wallet information
2. INJECT
It contains functionality for injecting code into other processes.
3. PERSIST
It can establish mechanisms that allow the malware to return after a reboot or continue running.
Researchers identified techniques involving:
Registry Run keys
Scheduled Tasks
Windows Management Instrumentation (WMI)
4. MOVE
The malware includes a decision option associated with lateral movement.
However, there is an important limitation:
The analyzed public build did not contain an operational handler for the “move” action.
So we should not portray CLOSEDQUORUM as an all-powerful autonomous worm.
It isn't.
But its architecture is what makes it important. (The Hacker News)
HERE'S THE REALLY INTERESTING PART
Traditional malware often depends on an attacker-controlled command-and-control infrastructure.
That creates something defenders can potentially investigate, block or disrupt.
CLOSEDQUORUM experiments with a different model:
The malware communicates with legitimate commercial AI services.
Gemini.
DeepSeek.
Qwen.
Mistral.
Those services are not inherently malicious.
They're legitimate AI platforms used by millions of people and organizations.
That creates a fascinating defensive problem.
Blocking every connection to those services simply because malware might abuse them isn't practical for many organizations.
Instead, defenders have to look at behavior.
For example:
Why is this Windows process communicating with multiple AI providers while simultaneously accessing LSASS, injecting into processes or creating persistence mechanisms?
That combination is far more suspicious than simply seeing traffic to an AI service.
THE MALWARE DOESN'T HAVE TO BE SMARTER THAN YOU
This is an important point.
CLOSEDQUORUM isn't some magical superintelligence.
The AI models aren't given unlimited freedom.
The malware restricts them to predefined choices.
And the system itself has weaknesses.
It depends on:
Internet connectivity
AI API availability
API credentials
Correct model responses
External services
The malware's own implementation
If the AI services refuse requests, return unusable responses, hit rate limits or become unavailable, the system can fail or wait and retry.
And according to Talos, the publicly distributed sample they analyzed contained placeholder API credentials and a dummy Discord webhook.
So the researchers did not observe the complete end-to-end autonomous attack cycle operating against a victim.
That distinction matters.
CLOSEDQUORUM is a demonstrated architectural concept not proof that criminals are currently deploying fully autonomous AI malware at scale. (blog.talosintelligence.com)
WHY SHOULD NIGERIAN BUSINESSES CARE?
Because the technology doesn't have to be sophisticated to become dangerous.
Many Nigerian businesses still rely heavily on:
Windows PCs
Internet banking
Email
Browser-stored passwords
Cryptocurrency wallets
Shared files
Remote-access tools
Cloud services
Imagine an employee's Windows computer becoming infected.
The attacker may not need to sit behind the keyboard continuously.
A malware system capable of making some tactical decisions could potentially continue operating while the human attacker is offline.
That means the old assumption:
“Someone has to be controlling the malware right now.”
may become increasingly unreliable.
YOUR ANTIVIRUS IS NOT THE WHOLE ANSWER
CLOSEDQUORUM also reinforces a broader cybersecurity lesson:
Detection has to become behavioral.
Don't only ask:
“Is this file malicious?”
Ask:
“What is this computer doing?”
A legitimate application suddenly:
Accessing LSASS
Reading browser credentials
Creating persistence
Injecting code into another process
Contacting several AI providers
Sending information through a Discord webhook
should immediately deserve attention.
Cisco Talos specifically recommends looking for combinations of suspicious behavior rather than simply blocking legitimate AI-provider domains. (The Hacker News)
WHAT SHOULD ORGANIZATIONS DO?
1. Keep Windows patched
Many attacks still begin with ordinary weaknesses.
2. Don't allow unnecessary administrator privileges
If malware lands on a machine, excessive privileges can dramatically increase the damage.
3. Monitor unusual process behavior
Security teams should watch for credential dumping, process injection and unexpected persistence.
4. Monitor AI-related network traffic
AI traffic isn't automatically suspicious.
But AI traffic combined with suspicious system activity deserves investigation.
5. Protect browser credentials
Don't treat saved passwords inside browsers as harmless.
6. Use MFA
Especially for email, financial systems, cloud accounts and administrator accounts.
7. Segment important systems
If one workstation is compromised, it shouldn't automatically provide a path to everything else.
8. Maintain offline or separately protected backups
Because today's malware doesn't necessarily have to behave like yesterday's malware.
THE BIGGER STORY ISN'T CLOSEDQUORUM
The biggest story isn't that one malware sample can ask four AI models what to do.
The bigger story is the direction of travel.
First, attackers used AI to help humans.
Then AI began helping with reconnaissance.
Then researchers demonstrated AI systems capable of performing portions of cyberattacks.
Now we're seeing malware architecture in which AI models participate directly in operational decisions.
The progression looks something like this:
The question isn't whether every piece of malware will suddenly become autonomous.
It won't.
The question is:
How much of an attack can eventually happen without a human making every decision?
And that is a cybersecurity question we should be asking now not after the technology becomes commonplace.
THE NIGERIAN CYBERSECURITY LESSON
For businesses, schools, organizations and individuals in Nigeria, this should reinforce one simple principle:
Don't build your security strategy around yesterday's threats.
A strong password is good.
Antivirus is good.
MFA is good.
Backups are good.
Updates are good.
But cybersecurity is becoming increasingly about how systems behave, not just what files they contain.
The threat landscape is changing.
Our defensive thinking has to change with it.
OSMALLAMINTECH TAKEAWAY
CLOSEDQUORUM is not evidence that we now have unstoppable AI malware.
It isn't.
It has technical limitations, depends on external AI services, and researchers have not confirmed successful real-world deployment.
But it demonstrates something important:
AI models can increasingly become components of cyberattack infrastructure.
And once that happens, the attacker doesn't necessarily need to manually decide every next move.
The malware can start asking the machines.
OSMALLAMINTECH QUESTION
If malware can eventually decide what to steal, where to move and how to persist without waiting for a human command...
At what point does malware stop being a tool and start becoming an attacker?
What do you think?
OSMALLAMINTECH
Cybersecurity Awareness | Threat Intelligence | Digital Resilience
This gives us a strong continuation of the AI hacker → AI agent → AI-driven malware storyline without simply recycling the previous articles. The key factual caveat is that Talos has not confirmed successful in-the-wild deployment, so the article keeps that distinction clear. (blog.talosintelligence.com)
Comments
Post a Comment