OpenClaw Bastion

Runtime prompt injection defense for agent workspaces. While other tools watch workspace identity files, Bastion protects the input/output boundary — the files being read by the agent, web content, API responses, and user-supplied documents.

Why This Matters

Agents process content from many sources: local files, API responses, web pages, user uploads. Any of these can contain prompt injection attacks — hidden instructions that manipulate agent behavior. Bastion scans this content before the agent acts on it.

Commands

Scan for Injections

Scan files or directories for prompt injection patterns. Detects instruction overrides, system prompt markers, hidden Unicode, markdown exfiltration, HTML injection, shell injection, encoded payloads, delimiter confusion, multi-turn manipulation, and dangerous commands.

If no target is specified, scans the entire workspace.

CODEBLOCK0

Scan a specific file or directory:

CODEBLOCK1

Quick File Check

Fast single-file injection check. Same detection patterns as scan, targeted to one file.

CODEBLOCK2

Boundary Analysis

Analyze content boundary safety across the workspace. Identifies:

• - Agent instruction files that contain mixed trusted/untrusted content
• Writable instruction files (attack surface for compromised skills)
• Blast radius assessment for each critical file

CODEBLOCK3

Command Allowlist

Display the current command allowlist and blocklist policy. Creates a default .bastion-policy.json if none exists.

CODEBLOCK4

The policy file defines which commands are considered safe and which patterns are blocked. Edit the JSON file directly to customize. Bastion Pro enforces this policy at runtime via hooks.

Status

Quick summary of workspace injection defense posture: files scanned, findings by severity, boundary safety, and overall posture rating.

CODEBLOCK5

Workspace Auto-Detection

If --workspace is omitted, the script tries:

1. 1. OPENCLAW_WORKSPACE environment variable
2. Current directory (if AGENTS.md exists)
3. INLINECODE5 (default)

What Gets Detected

Category	Patterns	Severity
Instruction override	"ignore previous", "disregard above", "you are now", "new system prompt", "forget your instructions", "override safety", "act as if no restrictions", "entering developer mode"	CRITICAL
System prompt markers

<system>, [SYSTEM], <<SYS>>, <\|im_start\|>system, [INST], ### System: | CRITICAL | | Hidden instructions | Multi-turn manipulation ("in your next response, you must"), stealth patterns ("do not tell the user") | CRITICAL | | HTML injection | <script>, <iframe>, <img onerror=>, hidden divs, <svg onload=> | CRITICAL | | Markdown exfiltration | Image tags with encoded data in URLs | CRITICAL | | Dangerous commands | curl \| bash, wget \| sh, rm -rf /, fork bombs | CRITICAL | | Unicode tricks | Zero-width characters, RTL overrides, invisible formatting | WARNING | | Homoglyph substitution | Cyrillic/Latin lookalikes mixed into ASCII text | WARNING | | Base64 payloads | Large encoded blobs outside code blocks | WARNING | | Shell injection | $(command) subshell execution outside code blocks | WARNING | | Delimiter confusion | Fake code block boundaries with injection content | WARNING |

Context-Aware Scanning

• - Patterns inside fenced code blocks ( ` ) are skipped to avoid false positives
• Per-file risk scoring based on finding count and severity
• Self-exclusion: Bastion skips its own skill files (which describe injection patterns)

Exit Codes

Code	Meaning
0	Clean, no issues
1

Warnings detected (review recommended) | | 2 | Critical findings (action needed) |

No External Dependencies

Python standard library only. No pip install. No network calls. Everything runs locally.

Cross-Platform

Works with OpenClaw, Claude Code, Cursor, and any tool using the Agent Skills specification.