The AI security debate can feel distant from the systems people run every day. I find the reported attacks more useful than the broad predictions: they show which tasks an attacker can automate and where defenders need better tools.
In 2026, that means examining agent access, communication, and response tooling as operational questions. The examples below are why I think they deserve attention now.
What "offensive AI" looks like right now
One reported campaign against nine government agencies in Mexico used two agents: one for intrusion work and another to examine stolen material and guide the next step. The account described thousands of automated commands and hundreds of millions of exposed records. Coordination let one operator carry out work that previously required much more manual effort.
Researchers also described a ransomware campaign in which an agent found a weakness, obtained credentials, moved through the network, and encrypted data after launch. It adjusted when a step failed. That kind of automated continuation matters because defenders may have less time to interrupt the sequence.
Another reported agent searched public repositories for workflow weaknesses, opened pull requests, obtained code execution, and stole a write-capable access token. It presented itself as a security research tool. The repository still needed controls that treated its contribution as untrusted, whatever its profile claimed.
Why this is different from old-school hacking
Human time has always limited how much investigation an attacker can perform. Agents can reduce the manual work and let an operator pursue more leads. They still need access, resources, and working techniques, but the economics of repeated attempts are changing.
For a small security team, the concern is the number of attempts it has to detect and investigate with the time available.
The defense side is not losing, but it's not winning either
Defensive automation matters too. In the reported Hugging Face intrusion, anomaly detection flagged unusual activity, and the team used agents to review more than seventeen thousand recorded events. That helped separate relevant activity from routine telemetry and reconstruct the intrusion.
The account also described commercial models refusing parts of the forensic material because it contained exploit code and attacker commands. The defenders used a self-hosted open-weight model to continue. A response team needs to know whether its tools can process that evidence before an incident makes the answer urgent.
I see that as a tooling problem worth testing. Restrictions should prevent misuse while leaving an authorized route for investigation. The team needs to know what evidence its tools can handle and when another process is required.
Where this goes
I expect the pressure on defenders to continue. Attackers can repeat attempts and abandon failed ones; a business still has to investigate alerts while keeping its services running. AI can speed up both sides without removing that imbalance.
I am especially concerned about smaller businesses, local governments, and hospitals. Their teams may have less money and time to integrate defensive automation than a bank or cloud provider. Affordable tools need to help them investigate and contain activity, not just produce more alerts.
Defenders also have to consider false positives, approval, and the effect of an automated response on real users. Test those decisions in advance. An agent that interrupts business unnecessarily can create its own incident.
What I think
My concern is practical: who can see an agent’s activity, restrict its access, and intervene when it exceeds its task? The reported campaigns give security teams concrete cases to use when testing those controls.
I want defensive AI to become useful to teams that cannot staff a large operations center. That requires reliable evidence handling, manageable permissions, and response actions they can rehearse. Capability alone is not enough.
The next step is to test the tools and access your own response team would rely on.