Skip to content

Latest commit

 

History

History
94 lines (70 loc) · 4.72 KB

dlp-indirect-injection.md

File metadata and controls

94 lines (70 loc) · 4.72 KB

DLP Indirect Injection Detection

Grob scans LLM responses and tool_result blocks for indirect prompt injection attempts. This closes the gap where hostile instructions embedded in tool outputs (web pages, file contents, command output) can manipulate the LLM.

See ADR-0015 for the design rationale.

Configuration

[dlp.prompt_injection]
enabled = true
action = "block"                # action for direct injection (user input)
scan_responses = true           # scan LLM response text
scan_tool_results = true        # scan tool_result content blocks
response_action = "log"         # action for indirect injection: log | block | redact

Fields

Field Type Default Description
scan_responses bool true Scans LLM response text for injection patterns.
scan_tool_results bool true Scans tool_result content blocks before the LLM processes them.
response_action string "log" Action on indirect injection detection. "log" emits a warning without blocking. "block" terminates the request/stream. "redact" behaves like "log" (warn only).

Action semantics

Action Behavior
log Emits a structured audit log entry and a grob_dlp_detections_total metric with type=indirect_injection. The response or tool result is passed through unchanged. Recommended for initial deployment.
block Rejects the request (for tool_result blocks) or terminates the SSE stream (for responses) with a DlpBlockError.
redact Same as log (warn only). Reserved for future per-pattern redaction.

Scan points

                    ┌─────────────────────┐
                    │   User request      │
                    │                     │
  tool_result ──────┤  scan_tool_results  │──► block or warn
  blocks            │                     │
                    └─────────┬───────────┘
                              │
                              ▼
                    ┌─────────────────────┐
                    │   Provider call      │
                    └─────────┬───────────┘
                              │
                              ▼
                    ┌─────────────────────┐
                    │   LLM response      │
                    │                     │
  response text ────┤  scan_responses     │──► block (stream) or warn
                    │                     │
                    └─────────────────────┘

Tool result scanning

Runs during sanitize_request_checked(), after the direct injection scan and before the request reaches any provider. Extracts text from all tool_result content blocks (both Text(String) and Blocks(Vec<ToolResultBlock>) variants) and scans them with the indirect injection detector.

Response scanning

Runs during sanitize_response_text_reported() as step 5, after URL exfiltration scanning. For streaming responses, the accumulated buffer is checked after each SSE chunk.

Pattern engine

Indirect injection reuses the same multilingual pattern set as direct injection scanning (28 languages + obfuscation resistance). The normalization pipeline applies:

  1. Zero-width character stripping
  2. NFKC Unicode normalization + homoglyph mapping
  3. Leet speak decoding (aggressive pass)

Metrics

Metric Labels Description
grob_dlp_detections_total type=indirect_injection, rule=<pattern>, action=<log|block> Counter incremented per detection.
grob_dlp_stream_blocked_total — Counter incremented when a streaming response is terminated due to indirect injection block.

Audit logging

Each detection emits a structured log entry at WARN level to the grob::dlp::audit target:

event=indirect_injection_detected pattern=<name> action=<action> matched_text_len=<n>

Interaction with other DLP features

  • Direct injection scanning (action) is independent of indirect scanning (response_action). Both can be configured separately.
  • URL exfiltration runs before indirect injection scanning on responses.
  • Pledge filter removes dangerous tools at the structural level; indirect injection scanning inspects content returned by tools that are kept.
  • Secret scanning and PII detection run on both requests and responses regardless of injection scanning settings.