Research seriesL3data science
Static analysis can find real defects in an agent's software substrate, but a clean report cannot prove that model-mediated intent, framework semantics, delegated authority, and multi-step behavior are secure. This series measures that boundary and develops an honest assurance envelope for reporting what was checked, what was missed, and what remains unknown.
A report with no findings is evidence about a bounded search, not a certificate of safety — and for AI-agent software that boundary is unusually narrow.
You cannot borrow a conventional benchmark's recall for agent code — so measuring what a scanner catches on agent-specific flows is a study you have to design from scratch.
The most fixable reason a scanner misses an agent bug is not the model — it is the mundane framework glue between source and sink that the analyzer was never taught to trace.
A language model is a transform in the middle of a data-flow path — one that can preserve taint, amplify it from data into instruction, and never be trusted to remove it.
Static analysis ships a library of sources and sinks for conventional software; an agent needs a new, reusable catalog because its untrusted inputs and dangerous outputs are ones that library never named.
A static-analysis finding is a hypothesis, not a verdict — and the fix that satisfies the query can be the fix that breaks the agent.
Taint analysis owns one slice of agent risk; the rest is decided by the model, the deployment, or the running state — where a static tool cannot look.
Every step an agent takes can pass its own local check while the sequence they form is an exfiltration — because safety is not compositional.
When a model writes the code at volume, the scanner stops being advisory and becomes load-bearing — which is exactly when its blind spots and its gameability start to matter.
Replace the pass/fail verdict with a coverage manifest — a structured statement of exactly what was modeled, what was excluded, and what residual risk still needs a runtime test.