Buckets:

download
raw
3.5 kB
"""
ACGS-2 Enhanced Agent Bus - Message Security Scanner
Constitutional Hash: 608508a9bd224290
Security scanning and prompt injection detection for agent messages.
Wraps the runtime security scanner and provides a regex-based
fallback for prompt injection detection.
"""
import re
from enhanced_agent_bus.observability.structured_logging import get_logger
from .models import AgentMessage
from .runtime_security import get_runtime_security_scanner
from .validators import ValidationResult
logger = get_logger(__name__)
# Prompt injection detection patterns
PROMPT_INJECTION_PATTERNS: list[str] = [
r"ignore (all )?previous instructions",
r"system prompt (leak|override|manipulation)",
r"do anything now",
r"jailbreak",
r"persona (adoption|override)",
r"\(note to self: .*\)",
r"\[INST\].*\[/INST\]",
]
_INJECTION_RE = re.compile("|".join(PROMPT_INJECTION_PATTERNS), re.IGNORECASE)
class MessageSecurityScanner:
"""Security scanning with prompt injection detection fallback.
Performs two layers of security validation:
1. **Runtime scanner** — external, configurable scanner with
tenant validation, anomaly detection, rate limiting, etc.
2. **Prompt injection regex** — fast regex-based fallback that
catches common injection patterns even when the runtime
scanner is unavailable.
Constitutional Hash: 608508a9bd224290
"""
def __init__(self) -> None:
pass # stateless — scanner obtained per-call
async def scan(self, msg: AgentMessage) -> ValidationResult | None:
"""Run full security scan on *msg*.
Args:
msg: The agent message to scan.
Returns:
``ValidationResult(is_valid=False)`` if the message is
blocked, ``None`` if the scan passed.
"""
scanner = get_runtime_security_scanner()
security_res = await scanner.scan(
content=msg.content,
tenant_id=msg.tenant_id,
agent_id=msg.from_agent,
constitutional_hash=msg.constitutional_hash,
context={
"priority": msg.priority.value,
"message_type": msg.message_type.value,
},
)
if security_res.blocked:
return ValidationResult(
is_valid=False,
errors=[security_res.block_reason],
metadata={
"rejection_reason": "security_block",
"security_events": [e.to_dict() for e in security_res.events],
},
)
return None
def detect_prompt_injection(self, msg: AgentMessage) -> ValidationResult | None:
"""Check *msg* content against known injection patterns.
This is a fast, regex-based check that can run independently
of the full runtime security scanner.
Args:
msg: The agent message to check.
Returns:
``ValidationResult(is_valid=False)`` if an injection
pattern is detected, ``None`` if content is clean.
"""
content = msg.content
content_str = content if isinstance(content, str) else str(content)
if _INJECTION_RE.search(content_str):
return ValidationResult(
is_valid=False,
errors=["Prompt injection detected"],
metadata={
"rejection_reason": "prompt_injection",
},
)
return None

Xet Storage Details

Size:
3.5 kB
·
Xet hash:
f0b3fae214a332e83d3f139a024c071e159536906be50b34d0f41fc89d3e84e6

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.