New Technique Medium TD-2026-0033 For AI Eng, AppSec, Red Team AI-assisted
Baya: compile-then-execute web agent resists prompt-injection hijack in WASP tests
Baya is a web agent that compiles the user's request into a Python program calling a Trusted Website API before touching the live site. The planning model never observes live webpage content, so untrusted data can only set values or pick among pre-defined branches. On the WASP security benchmark, Baya reports a 0% attack success rate versus 16.2% for API-enabled ReAct and 25.5% for browser-only ReAct.
Why it matters
Organizations deploying web agents face prompt injection from webpages that can hijack the agent's control flow. This design aims to remove that path by fixing the task plan before live content is read, though the results come from a single paper's evaluation.
Do this
- Treat the reported 0% WASP attack success rate as a research result. Validate it on your own agent workloads before relying on it.
- For web agents in production, limit which actions can run after reading untrusted page content, and require approval for state-changing operations.
- Where a trusted, structured API for a website exists, prefer API-scoped actions over free-form browser actions to reduce the injectable surface.
- Log and review agent tool calls that deviate from the originally approved task plan to detect control-flow hijacking attempts.
Signals
- NoCISA KEV
- 23/100Priority
- 60%Confidence
Evidence
“Before accessing the live website, Baya compiles the user's request into an executable Python program that calls a Trusted Website API.”
“On WASP, a security benchmark for web agents, Baya achieves 0% attack success rate, compared with 16.2% for API-enabled ReAct and 25.5% for browser-only ReAct.”
Sources
Affected / related
compile-then-execute web agent, Baya, ReAct paradigm, WebArena Verified, WASP
Published 2026-11-04 06:42 UTC · updated 2026-11-08 17:52 UTC · AI-assisted summary, checked against quoted sources · How we score