script injection

Summary

Premise

• commands/queries are composedby concatenating or untrusted input into executable context (SQL statements, shell commands, prompts)
• Server executes resulting code/query with elevated privileges

Attack

• SQL injection: attacker injects SQL fragments to change the query
• Command injection: inject malicious commands into shell-invoked strings
• Prompt injection (AI): craft input that alters/overrides model instructions

Attacker’s goals

• access backend data/read files -> confidentiality
• modify backend data -> integrity
• execute arbitrary/manipulate AI outputs -> execution integrity

Defense

• use parameterized queries, avoid string concatenation
• avoid invoking shell
• validate and sanitize inputs, whitelisting
• for AI systems: sanitize prompts, use model guards, restrict what models can access, and apply output-filtering
• principle of least privilege for DB/accounts and RBAC for commands

Concept

Script injection

• no separation between data and script
• easy for data to get interpreted as script

SQL injection

• database query

SELECT * FROM users WHERE name = 'Bob'
-- standard query that searches for the user Bob

SELECT * FROM users WHERE name = '$userinput' 
-- query constructed by login page, checks if the user inputted name exists

-- "with any' OR 1=1 --"
SELECT * FROM users WHERE name = 'any' OR 1=1 --' 
-- leaks the whole users table, OR 1=1 gets evaluated as a script instead of a string

Input validation

• filtering
• whitelist - list of items that are known to be benign and accepted
• blacklist - list of items known to be bad and rejected
• requirements:
  • complete - doesn’t miss out any malicious string
  • accepts all legitimate inputs

Parameterized queries

• clear separation between data and script
• parameters are sent separately from the rest of the query