The architecture changed. The ownership model did not.
For AI security engineers hired to secure agents and the systems they touch. The work cuts across AppSec, IAM, data, and governance. The code is where those duties meet.
Cloud security used to be treated as another responsibility for the existing team. Then the architecture changed enough that the old ownership model stopped scaling, and the role became its own function. AI is at that point.
An agent can reason over data, choose a tool, inherit permissions, and take an action in another system. Agent identity, prompt paths, RAG and MCP wiring, sandboxing, and runtime containment now show up in job descriptions at companies that used to fold this into AppSec. The organizational gap those roles fill is ownership of the whole chain. The code-review gap is narrower: the permissions, logs, and human-oversight paths that make the chain real are decided in pull requests.
The chain is organizational. The diff is still a control.
Runtime tools watch the agent. heygrc reads the change that gave it permission. Linters and SAST are not built to know which frameworks you carry. A prompt-injection test does not see that last week's PR deleted the override path or widened the tool-calling role. heygrc is built to read each pull request against the frameworks you selected and cite the control a change touches: agent identity and access, record-keeping over the system's operation, the human's way to intervene.
Art. 12 and Art. 14 are high-risk duties. heygrc does not classify your system's risk tier or run your conformity assessment. Whether those articles apply is your own assessment. It does not sandbox the agent, test for prompt injection, or contain a runaway tool call at runtime. It reviews the code change against the frameworks you selected and cites the control. You stay the judge of what blocks and what ships.
Changes that read as ordinary code.
A few of the control-relevant changes heygrc is built to flag for this case, each cited to the clause it touches.
An agent or tool-calling role widens to a wildcard
SOC 2 CC6.1The human override is removed from the agent reviewer console
EU AI Act Art. 14Inference logging is removed from a decision-making agent
EU AI Act Art. 12
The frameworks that matter most here.
Guide: EU AI Act for developers
heygrc flags control-relevant changes and cites the clause so the issue can be handled in the pull request. It does not certify you, run your audit, or replace your own judgment.