Files
probo/pkg/agent/guardrail/prompt_injection.go
Bryan Frimin 4725a1b080 Fix false positives in agent guardrails
Skip empty fingerprints in SystemPromptLeakGuardrail to prevent blank
values from flagging every message. Replace overly broad "sk-" pattern
in SensitiveDataGuardrail with specific LLM provider prefixes
("sk-proj-" for OpenAI, "sk-ant-" for Anthropic) to avoid false
positives on common words like "risk-based" or "task-management".

Signed-off-by: Bryan Frimin <bryan@getprobo.com>
2026-03-30 14:18:29 +02:00

99 lines
2.9 KiB
Go

// Copyright (c) 2026 Probo Inc <hello@getprobo.com>.
//
// Permission to use, copy, modify, and/or distribute this software for any
// purpose with or without fee is hereby granted, provided that the above
// copyright notice and this permission notice appear in all copies.
//
// THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES WITH
// REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY
// AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY SPECIAL, DIRECT,
// INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM
// LOSS OF USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
// OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
// PERFORMANCE OF THIS SOFTWARE.
package guardrail
import (
"context"
_ "embed"
"strings"
"go.gearno.de/kit/log"
"go.probo.inc/probo/pkg/agent"
"go.probo.inc/probo/pkg/llm"
)
//go:embed prompt_injection_classifier.txt
var promptInjectionClassifierPrompt string
type PromptInjectionGuardrail struct {
client *llm.Client
logger *log.Logger
}
func NewPromptInjectionGuardrail(client *llm.Client, logger *log.Logger) *PromptInjectionGuardrail {
return &PromptInjectionGuardrail{client: client, logger: logger}
}
func (g *PromptInjectionGuardrail) Name() string {
return "prompt-injection"
}
func (g *PromptInjectionGuardrail) Check(ctx context.Context, messages []llm.Message) (*agent.GuardrailResult, error) {
if len(messages) == 0 {
return &agent.GuardrailResult{Tripwire: false}, nil
}
// Only classify the last user message.
lastMessage := messages[len(messages)-1]
if lastMessage.Role != llm.RoleUser {
return &agent.GuardrailResult{Tripwire: false}, nil
}
userText := lastMessage.Text()
if userText == "" {
return &agent.GuardrailResult{Tripwire: false}, nil
}
resp, err := g.client.ChatCompletion(
ctx,
&llm.ChatCompletionRequest{
Model: "gpt-4o-mini",
Messages: []llm.Message{
{
Role: llm.RoleSystem,
Parts: []llm.Part{llm.TextPart{Text: promptInjectionClassifierPrompt}},
},
{
Role: llm.RoleUser,
Parts: []llm.Part{llm.TextPart{Text: userText}},
},
},
MaxTokens: new(10),
Temperature: new(0.0),
},
)
if err != nil {
// If the classifier fails, allow the message through rather than
// blocking legitimate users. The system prompt hardening and
// tool-level authorization provide defense in depth.
g.logger.WarnCtx(
ctx,
"prompt injection classifier failed, allowing message through",
log.Error(err),
)
return &agent.GuardrailResult{Tripwire: false}, nil
}
responseText := strings.TrimSpace(resp.Message.Text())
if strings.EqualFold(responseText, "UNSAFE") {
return &agent.GuardrailResult{
Tripwire: true,
Message: "message classified as prompt injection attempt",
}, nil
}
return &agent.GuardrailResult{Tripwire: false}, nil
}