A practice built around one problem
GambitSec works on adversarial robustness and cost-governed autonomy for cyber-defense agents: AI systems that read security telemetry and are allowed to act on it.
The input is the attack surface
Security teams are handing alert triage and first response to LLM agents because alert volume has outgrown the people available to read it. Those agents are fed the same logs, alerts and tickets an analyst would read, and a large share of that text was written by whoever caused the alert.
That makes a defensive agent different from most AI deployments. Its adversary does not need access to the model or the prompt. They only need to put text somewhere the pipeline will pick it up, and the agent may have authority to close alerts, isolate hosts or disable accounts based on what it reads.
The second constraint is cost. Every check added to make an agent safer spends tokens, and alert volume is partly under the attacker's control. A design that ignores either side fails in production: it is unsafe to automate or too expensive to keep running.