Files
agents_and_robots/pkg/sanitize/sanitize.go
T
egutierrez e8dd7c41ed feat: crear pkg/sanitize para deteccion de prompt injection
Nuevo paquete puro (sin I/O) que detecta patrones de prompt injection
en mensajes de usuario antes de enviarlos al LLM.

- patterns.go: 15 patrones en ingles y español (delimitadores de sistema,
  override de instrucciones, exfiltracion de prompt, jailbreak, evasion base64)
- sanitize.go: funcion Sanitize() con 3 modos (warn, strip, reject),
  filtro por severidad minima y patrones deshabilitables
- Tipos: Pattern, Severity, Mode, Options, Warning, Result

Todo puro: string in → Result out. Los side effects (logging, rechazo)
ocurren en el caller (runtime.go).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-03-07 19:34:24 +00:00

137 lines
3.1 KiB
Go

package sanitize
import "strings"
// Mode controls how the sanitizer handles detected patterns.
type Mode int
const (
ModeWarn Mode = iota // report warnings but don't modify the message
ModeStrip // remove matched patterns from the message
ModeReject // reject the message entirely if any pattern matches
)
func (m Mode) String() string {
switch m {
case ModeWarn:
return "warn"
case ModeStrip:
return "strip"
case ModeReject:
return "reject"
default:
return "unknown"
}
}
// ParseMode converts a string to a Mode. Returns ModeWarn for unrecognized values.
func ParseMode(s string) Mode {
switch strings.ToLower(s) {
case "strip":
return ModeStrip
case "reject":
return ModeReject
default:
return ModeWarn
}
}
// Options configures the sanitizer behavior.
type Options struct {
Mode Mode // how to handle detections
MinSeverity Severity // only act on patterns at or above this severity
Patterns []Pattern // patterns to check (nil = DefaultPatterns)
DisabledPatterns []string // pattern names to skip
}
// Warning represents a detected prompt injection pattern in the input.
type Warning struct {
PatternName string // which pattern matched
Severity Severity // threat level
Matched string // the text that matched (first match only)
}
// Result holds the output of a Sanitize call.
type Result struct {
Output string // the (possibly modified) message
Warnings []Warning // all detected patterns
Rejected bool // true if the message was rejected (ModeReject + match found)
}
// Sanitize checks the input for prompt injection patterns and returns
// the result according to the configured mode.
//
// This is a pure function: no I/O, no side effects.
func Sanitize(input string, opts Options) Result {
patterns := opts.Patterns
if patterns == nil {
patterns = DefaultPatterns()
}
disabled := make(map[string]bool, len(opts.DisabledPatterns))
for _, name := range opts.DisabledPatterns {
disabled[name] = true
}
var warnings []Warning
output := input
for _, p := range patterns {
if disabled[p.Name] {
continue
}
if p.Severity < opts.MinSeverity {
continue
}
loc := p.Regex.FindStringIndex(output)
if loc == nil {
continue
}
matched := output[loc[0]:loc[1]]
warnings = append(warnings, Warning{
PatternName: p.Name,
Severity: p.Severity,
Matched: matched,
})
if opts.Mode == ModeStrip {
output = p.Regex.ReplaceAllString(output, "")
}
}
result := Result{
Output: output,
Warnings: warnings,
}
if opts.Mode == ModeReject && len(warnings) > 0 {
result.Rejected = true
}
return result
}
// HasHighSeverity returns true if any warning is SeverityHigh.
func (r Result) HasHighSeverity() bool {
for _, w := range r.Warnings {
if w.Severity == SeverityHigh {
return true
}
}
return false
}
// MaxSeverity returns the highest severity among all warnings.
// Returns SeverityLow if there are no warnings.
func (r Result) MaxSeverity() Severity {
max := SeverityLow
for _, w := range r.Warnings {
if w.Severity > max {
max = w.Severity
}
}
return max
}