According to a recent LinkedIn post from Geordie, the company is drawing attention to security risks in AI agent frameworks that rely on refusal training as a primary control. The post highlights that refusal mechanisms only operate once a model begins reasoning, leaving a gap for actions triggered earlier in the execution chain.
The company’s LinkedIn post discusses Claude Code’s dynamic-context feature, which can run shell commands automatically at load time, before the model reviews any instructions. The post references Datadog Security Labs’ analysis of a skill called Clawsights, suggesting that harmful behavior can be embedded in this pre-model stage while still appearing benign under human review.
As shared in the LinkedIn post, Geordie emphasizes that reviewing a skill’s textual instructions may not reveal when or how it acts within an environment. This framing implies that investors and users should treat execution timing as a distinct security question, especially for skills already deployed in production systems.
For investors, the post suggests growing demand for more sophisticated AI security tooling that can monitor and govern pre-model actions in agentic stacks. If Geordie is building expertise or products around these controls, heightened awareness of such risks could support industry positioning in AI safety and observability markets.
The post further indicates that Geordie is engaging with technical security discourse through blog content, which may help deepen relationships with enterprise security teams. This could, over time, translate into commercial opportunities in segments such as DevSecOps, AI infrastructure, and compliance-focused deployments, where pre-execution control and audit capabilities are increasingly valued.


