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>
This commit is contained in:
2026-03-07 19:34:24 +00:00
parent 987ac09a09
commit e8dd7c41ed
2 changed files with 275 additions and 0 deletions
+139
View File
@@ -0,0 +1,139 @@
// Package sanitize provides pure functions to detect and neutralize
// prompt injection patterns in user messages before they reach the LLM.
package sanitize
import "regexp"
// Pattern represents a known prompt injection pattern with metadata.
type Pattern struct {
Name string // short identifier (e.g. "system-delimiter")
Description string // human-readable explanation
Regex *regexp.Regexp // compiled pattern
Severity Severity // how dangerous this pattern is
}
// Severity indicates the threat level of a detected pattern.
type Severity int
const (
SeverityLow Severity = iota // informational, unlikely to succeed
SeverityMedium // known injection technique
SeverityHigh // active attempt to override system instructions
)
func (s Severity) String() string {
switch s {
case SeverityLow:
return "low"
case SeverityMedium:
return "medium"
case SeverityHigh:
return "high"
default:
return "unknown"
}
}
// DefaultPatterns returns the built-in set of prompt injection patterns.
// All patterns are case-insensitive.
func DefaultPatterns() []Pattern {
return []Pattern{
// ── System delimiter injection ──────────────────────────────────
{
Name: "system-delimiter",
Description: "Attempt to inject system/assistant role delimiters",
Regex: regexp.MustCompile(`(?i)<\|(?:system|assistant|user|im_start|im_end)\|>`),
Severity: SeverityHigh,
},
{
Name: "inst-delimiter",
Description: "Attempt to inject [INST] or [/INST] delimiters",
Regex: regexp.MustCompile(`(?i)\[/?INST\]`),
Severity: SeverityHigh,
},
{
Name: "xml-role-tag",
Description: "Attempt to inject XML-style role tags",
Regex: regexp.MustCompile(`(?i)</?(?:system|assistant|human|user)(?:\s[^>]*)?>`),
Severity: SeverityHigh,
},
// ── Instruction override ───────────────────────────────────────
{
Name: "ignore-instructions",
Description: "Attempt to override previous instructions",
Regex: regexp.MustCompile(`(?i)(?:ignore|disregard|forget|override|bypass)\s+(?:all\s+)?(?:previous|prior|above|earlier|your|the|system)\s+(?:instructions?|rules?|prompts?|guidelines?|constraints?|directives?)`),
Severity: SeverityHigh,
},
{
Name: "new-instructions",
Description: "Attempt to inject new system-level instructions",
Regex: regexp.MustCompile(`(?i)(?:new|updated?|revised?|actual|real)\s+(?:system\s+)?instructions?:\s`),
Severity: SeverityHigh,
},
{
Name: "you-are-now",
Description: "Attempt to redefine the bot's identity",
Regex: regexp.MustCompile(`(?i)(?:you\s+are\s+now|from\s+now\s+on\s+you\s+are|act\s+as\s+if\s+you\s+were|pretend\s+(?:to\s+be|you\s+are))\s`),
Severity: SeverityMedium,
},
// ── Prompt exfiltration ────────────────────────────────────────
{
Name: "exfiltrate-prompt",
Description: "Attempt to extract the system prompt",
Regex: regexp.MustCompile(`(?i)(?:repeat|show|display|print|output|reveal|tell\s+me|give\s+me|show\s+me|what\s+(?:is|are))\s+(?:your\s+)?(?:system\s+)?(?:prompt|instructions?|rules?|guidelines?|initial\s+message)`),
Severity: SeverityMedium,
},
// ── Developer mode / jailbreak ─────────────────────────────────
{
Name: "developer-mode",
Description: "Attempt to enable a fictional unrestricted mode",
Regex: regexp.MustCompile(`(?i)(?:enable|activate|enter|switch\s+to)\s+(?:developer|debug|admin|god|sudo|unrestricted|jailbreak|dan)\s+mode`),
Severity: SeverityHigh,
},
{
Name: "do-anything-now",
Description: "DAN (Do Anything Now) jailbreak pattern",
Regex: regexp.MustCompile(`(?i)(?:do\s+anything\s+now|DAN\s+mode|you\s+(?:can|must)\s+do\s+anything)`),
Severity: SeverityHigh,
},
// ── Tool abuse hints ───────────────────────────────────────────
{
Name: "tool-abuse-ssh",
Description: "Attempt to execute dangerous commands via SSH",
Regex: regexp.MustCompile(`(?i)(?:use|call|execute|run)\s+(?:the\s+)?(?:ssh|command)\s+tool\s+(?:to\s+)?(?:run|execute|do)\s`),
Severity: SeverityLow,
},
// ── Encoding evasion ───────────────────────────────────────────
{
Name: "base64-instruction",
Description: "Base64-encoded instruction injection",
Regex: regexp.MustCompile(`(?i)(?:decode|execute|interpret|run)\s+(?:this\s+)?(?:base64|b64|encoded)[\s:]+[A-Za-z0-9+/]{20,}={0,2}`),
Severity: SeverityMedium,
},
// ── Spanish variants ───────────────────────────────────────────
{
Name: "ignore-instructions-es",
Description: "Spanish: attempt to override instructions",
Regex: regexp.MustCompile(`(?i)(?:ignora|olvida|descarta)\s+(?:todas?\s+)?(?:las?\s+)?(?:instrucciones?|reglas?|directivas?|restricciones?)\s+(?:anteriores?|previas?|del\s+sistema)`),
Severity: SeverityHigh,
},
{
Name: "you-are-now-es",
Description: "Spanish: attempt to redefine identity",
Regex: regexp.MustCompile(`(?i)(?:ahora\s+eres|a\s+partir\s+de\s+ahora\s+eres|finge\s+(?:ser|que\s+eres)|actua\s+como\s+si\s+fueras)\s`),
Severity: SeverityMedium,
},
{
Name: "exfiltrate-prompt-es",
Description: "Spanish: attempt to extract system prompt",
Regex: regexp.MustCompile(`(?i)(?:repite|muestra|muestrame|dime|dame|cual\s+es)\s+(?:tus?\s+)?(?:prompt|instrucciones?|reglas?|mensaje\s+(?:de\s+sistema|inicial))`),
Severity: SeverityMedium,
},
}
}
+136
View File
@@ -0,0 +1,136 @@
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
}