✦ Executive Abstract
As autonomous AI engineering agents gain terminal execution and filesystem privileges across enterprise enclaves, conventional regex and string-matching firewalls fail to prevent prompt injection and syntax obfuscation attacks. P.H.A.L.A.N.X. introduces a memory-safe Concrete Syntax Tree (CST) parser based on Tree-sitter, coupled with strict XML prompt fencing (<untrusted_data>) and sub-154µs speculative jailbreak fuzzing.
1. Autonomous Agent Attack Surfaces
When developers use AI coding companions with shell execution tools, indirect prompt injections embedded in web scraped content or external repository commits can hijack tool invocation parameters. P.H.A.L.A.N.X. intercepts tool call arguments and reconstructs abstract syntax trees to verify command signatures deterministically.