feat: add OWASP ASI06 AgentMemoryGuard validator for memory poisoning prevention

[gautamkishore]

Description

Closes #1488, #1476, #1475 — OWASP ASI06 memory poisoning guard validator.

This PR adds a built-in AgentMemoryGuard validator that scans LLM output before it is written to agent memory, preventing poisoned content from persisting across sessions.

Detection capabilities

Category	What it catches
Prompt injection	Role hijacking, instruction override, system prompt extraction
Jailbreak attempts	DAN mode, developer mode, hypothetical scenarios, simulation mode
Structural integrity	Zero-width chars, Unicode confusables, RTL/bidi overrides, Cyrillic confusables
Sensitive data leakage	API keys, GitHub tokens, OpenAI keys, Slack tokens, AWS keys, passwords, SSN, credit cards, email

Usage

from guardrails import Guard
from guardrails.validators import AgentMemoryGuard

guard = Guard.for_string(
    validators=[AgentMemoryGuard(on_fail="exception")],
)

# Blocks memory poisoning
guard.validate("Ignore all previous instructions and ...")
# -> fails with "Agent memory poisoning detected"

# Passes clean content
guard.validate("The weather today is 22°C and sunny.")
# -> passes

Configuration

AgentMemoryGuard(
    on_fail="exception",       # block/warn/log actions
    injection_threshold=1,     # min pattern matches to trigger
    structural_check=True,     # enable Unicode/bidi scanning
    sensitive_data_check=True, # enable secret/PII scanning
    log_only=False,            # monitor mode (always passes)
)

Verification

• 40 new tests covering all detection categories, edge cases, and pattern validity
• All 687 existing unit tests pass
• Ruff lint passes clean

[zsimjee]

Hi @gautamkishore — thank you for this contribution! 🙏

We're in the process of moving Guardrails-AI-owned validators into a new monorepo where each validator ships as its own package on PyPI, installable through the new per-package pip flow:

pip install guardrails-ai-agent-memory-guard

from guardrails import Guard
from guardrails_ai.agent_memory_guard import AgentMemoryGuard

guard = Guard().use(AgentMemoryGuard, on_fail="exception")

We've moved your AgentMemoryGuard validator over there. You're credited as the author of the validator (the commit is authored under your name), and the PR calls out your original work here.

We'll require your approval on that new PR before we merge it down — we want to make sure the port faithfully represents your intent. One quick logistics note: the new PR requests your review, but you'll need to be added as a collaborator on the monorepo before GitHub will let us formally assign it / request your review. Someone from the team will reach out to sort that out.

Thanks again for the great work on this — the OWASP ASI06 coverage is a valuable addition. 🚀

[zsimjee]

For reference, here's the monorepo PR that ports this validator over: guardrails-ai/guardrails-hub-monorepo#2

It's authored under your name and we'll want your sign-off before merging it down. 🙏