LLMs and sandbox primitives cut cost and boost security for web extensible software

The core change described is that large language models make it far cheaper to write web extensions, while modern sandboxing techniques reduce the cost and risk of deploying those extensions. Together they enable a model where an application’s core remains fixed and accountable, and users can safely add functionality through LLM-generated code. This represents a shift from traditional plug-in development, which often required significant engineering effort and careful security review.

Adopting this model entails integrating an LLM generation pipeline into the development workflow, configuring sandbox environments that enforce isolation, and establishing processes to review or validate LLM output before it is exposed to users. Teams must also invest in monitoring to detect hallucinated or unsafe behavior that could bypass sandbox limits. The core application must be designed with clear extension points and contracts so that LLM-filled pieces can be reliably composed.

The approach stops working when extensions need privileged system access, real-time guarantees, or deep integration with low-level components that sandbox primitives cannot safely isolate. It also falters if the LLM frequently produces incorrect or insecure code, requiring extensive manual correction that erodes the cost benefits. In domains where determinism and auditability are paramount, reliance on stochastic language models introduces unacceptable uncertainty.

Only one source (Simon Willison’s weblog) reports Jeremy Morrell’s hypothesis, so there is no independent corroboration at this time. Engineers should treat the idea as a promising but unverified concept, seeking further evidence or case studies before committing significant resources to an LLM-driven extensibility strategy.