While organizations build internal agent capabilities

The key shift is that AI is becoming a first pass operator inside the SOC, but not yet a trusted closer. In practice, the agent reads alert data in Splunk, opens or updates work in Jira, and can inspect code or repo context in GitHub, then hands back a recommended disposition for an analyst to approve. That lets teams remove repetitive tier 1 work without giving the model final authority over incident response.

• This setup follows the same control pattern already baked into security tooling. Splunk SOAR supports approval steps where a reviewer decides whether an action is allowed to run, which mirrors the internal workflow here of agent recommendation first, human execution second.
  1 sacra 4 splunk
• The initial use case is narrow for a reason. Duplicate alerts and obvious false positives are high volume tasks with lots of historical examples, so teams can compare the agent's suggestion against past analyst decisions before trusting it with harder calls like containment or remediation.
  1 sacra
• The broader market is moving the same way. GitHub's AI security features can suggest code fixes and review pull requests automatically, but GitHub explicitly says teams should validate the feedback and review suggested changes before accepting them. Vendor products like Sublime package more of this workflow, while large enterprises often build it themselves on top of existing systems.
  2 github 3 github 5 sacra

Over the next phase, the handoff point will move from recommendation to bounded execution. The first fully automated actions are likely to be low risk closures, like suppressing duplicate alerts or closing well understood false positives, while higher consequence actions stay gated behind analyst approval until teams build enough logged history to trust the model's judgment.