Opinion
AI is giving cyber defenders a home advantage
"Defenders have an advantage an attacker does not," writes Sweet Security CEO Dror Kashti. "They are already inside the environment they need to understand."
Most discussions about AI and cybersecurity begin from the same assumption: attackers will move faster, scale further and become more effective. That is true. But it is only half of the story.
Defenders have an advantage an attacker does not. They are already inside the environment they need to understand. They can see the applications running in production, the identities those applications use, the services they communicate with and the data moving between them. AI makes that advantage far more valuable - and far more practical to use.
The advantage begins with knowing the environment as it actually operates. A diagram describes how services are connected. A permissions policy describes what an identity may access. Runtime evidence records what software actually does while it is running. Connected with cloud activity and identity information, it shows which process used an identity, which service it contacted and what happened next. Those relationships give a security decision its context - and they belong to the defender, not the attacker.
For years, assembling that picture was expensive. Evidence sat in different tools. Reconstructing an incident meant moving between application activity, cloud logs, identities and infrastructure while the environment kept changing. AI can change this by reasoning continuously across those relationships - learning what an application normally does, identifying the resources it depends on and recognizing when a new sequence of actions creates risk. But AI is only as useful as the evidence beneath it. The strongest decisions come from direct, current evidence of how the environment is actually operating.
The same principle applies to penetration testing. An outside attacker begins with incomplete information. A defender with access to the codebase, configurations and runtime behavior can use this to find real and complex attack paths in applications and infrastructure, rather than evaluating each misconfiguration or vulnerability in isolation.
We saw this clearly with Sweet Attack, our AI red-team agent. A weakness in a customer-facing application allowed the agent to direct requests to an internal management tool. That tool accepted requests without authentication and could execute commands. The workload running those commands had an identity with broad cloud permissions. Individually, each condition told only part of the story. Together, they created a validated path from an unauthenticated external request to cloud administrator access - in under four minutes.
A validated attack path gives defenders a specific place to act. Fix the condition that made the path possible. Then enforce protection while the business is running.
That last step is critical. Security cannot end with an alert or a remediation ticket. If an AI agent with legitimate access is manipulated into sending sensitive data to an unauthorized destination, it must be stopped short. Logging or identifying such behaviors post-exploitation is important, but the damage is already done. This is why runtime enforcement is key.
This creates a continuous cycle: attack, fix and defend. Test the environment to prove which paths are exploitable. Fix the conditions that create real exposure. Defend the running environment when harmful behavior occurs. Then feed what happens in production back into the next round.
The cycle continues because production never stands still. A deployment changes a dependency. A new agent expands reach. A permission is added for a legitimate reason but creates an unintended path somewhere else. Effective enforcement requires precision: an understanding of intended use, current runtime behavior and proven attack paths - combined with the ability to stop the harmful action while allowing the business to keep moving.
Attackers will use AI to move faster. Defenders can use it with deeper context, better access and the authority to act. The organization that knows its environment as it runs, continuously tests how it can be exploited and enforces protection at runtime has the real home advantage.
Dror Kashti is co-founder and CEO of Sweet Security.
