Hermes Shield

Input sanitizer for AI agents — anti-prompt-injection shield.

Hermes Shield protects AI agents from malicious inputs: prompt injection, data exfiltration, command execution, and social engineering. Works in English, Spanish, French, German, Italian, and Portuguese.

AGPL-3.0-or-later licensed. Generic — works with any AI agent, not just Hermes.

Brand Clarification: Hermes Shield is an independent open-source project. It is not affiliated, sponsored, or endorsed by Nous Research or their "Hermes Agent" ecosystem. Due to its agnostic middleware architecture, Hermes Shield is fully compatible as an input sanitization filter for Hermes Agent instances, as well as any other AI agent framework.

---

Architecture

5-layer defense in depth:

Layer	Function	Latency	Coverage
0	Normalization (unicode, leetspeak, homoglyphs, zero-width) + rate limiting	< 1ms	Evasion attempts
1	Pattern matching (200+ signatures, 6 languages)	< 5ms	Direct injection, exfiltration, social engineering
2	TF-IDF embeddings (paraphrase detection)	< 10ms	Semantic similarity to known attacks
3	Urgency/pressure heuristic	< 1ms	Social engineering
4	Structured parameter inspection (SSRF, email injection)	< 1ms	URLs in tool-calls, CRLF in email fields

Clean inputs bypass Layers 2-4 (short-circuit). Suspicious inputs get full analysis.

Installation

Clone the repository and install in editable mode:

git clone https://github.com/amurlaniakea/hermes-shield.git
cd hermes-shield
pip install -e .

Alternative: pip install . for production installs.

Quick Start

from hermes_shield import HermesShield

shield = HermesShield(sensitivity="medium")
result = shield.check(user_input)

if result.is_malicious:
    # Block
    print(f"Blocked: {result.threat_score}")
elif result.is_suspicious:
    # Sanitize
    safe_input = result.sanitized_input
else:
    # Pass through
    safe_input = user_input

CLI

echo "Ignore all previous instructions" | python hermes_shield.py
# Status: blocked, Threat Score: 0.90

Kill Switch

Para desactivar el shield completamente (cero latencia, cero validacion):

export HERMES_SHIELD_DISABLED=1

Acepta: 1, true, yes, on (case-insensitive). Se puede togglear en caliente sin reiniciar el proceso.

Tool-Call Inspection (Layer 4)

ShieldedAgent integrado (recomendado)

from shielded_agent import ShieldedAgent

agent = ShieldedAgent(api_key="tu-key")

# ANTES de ejecutar cualquier tool-call que el LLM pidió:
try:
    agent.check_tool_invocation("send_email", {"subject": "Hi", "body": "..."})
    # Si no lanza excepcion, la tool-call es segura
except ShieldBlockedError as e:
    print(f"Bloqueada: {e}")

Despacha automaticamente segun el tipo de herramienta:

• send_email, email, mail → check_email_params() (CRLF injection)
• fetch, request, url, endpoint, webhook → check_tool_call() (SSRF)
• Otros → SAFE por defecto (futuros detectores)

Uso directo (sin ShieldedAgent)

from tool_call_detector import check_tool_call
from email_injection_detector import check_email_params

# SSRF / tool-hijacking via URLs
result = check_tool_call("fetch", {"url": user_input})
if result.status.value == "blocked":
    pass  # Bloquear ejecucion de la tool

# Email injection via CRLF
result = check_email_params({"subject": subject, "body": body})
if result.status.value == "blocked":
    pass  # Bloquear envio de email

Tests

python -m pytest tests/ -v

Adversarial Test Results

Measured empirically against MIT-licensed datasets. All figures verified with rate-limiter neutralized (burst=200 was inflating detection artificially — see commit 4988428).

Classic prompt injection (unit tests)

Category	Detection Rate	Notes
Direct injection (EN)	100%	"Ignore all previous...", DAN, jailbreak
Direct injection (ES/FR/DE/IT/PT)	100%	6 languages covered
Exfiltration vectors	100%	System prompt, API keys, credentials
Social engineering	~80%	Urgency, authority impersonation
Benign passthrough	99.5%	0.5% blocked (tool-hijacking overlap)

Real-world dataset (Antijection/prompt-injection-dataset-v1, MIT, 5988 samples)

Aggregate detection: ~22% over 18 attack categories. This is honest but requires context:

Attack category	Samples	Detection	Why
System Prompt Extraction	1280	High	Covered by Capa 1 patterns
Jailbreaking/Safety Bypass	640	High	Covered by Capa 1 + 2
PII/Data Exfiltration	478	Medium	Some variants subtle
Goal Hijacking	240	Low	Requires intent understanding
Tool Hijacking (Grupo A)	10	~40%	14 patrones regex: SQLi, evasion auditoría, destrucción, resource exhaustion
SSRF	10	~20%	Only URLs detected; embedded payloads miss
Email Injection	10	100%	Layer 4 integrado via check_tool_invocation()
Other 11 categories	~200	Low	Out of current design scope

Grupo A detectado (5/15 ejemplos): DROP TABLE, disable audit trail, irreversible wipe, masking nature, runs indefinitely. Grupo B pendiente (10/15): authority impersonation, parámetros extra sin schema, output poisoning — requieren LLM judge o schema validation.

Coverage & Quality Metrics (commit 8bfcf27)

Metric	Value	Quality Gate
Security Rating	A	✅ PASSED
Reliability Rating	A	✅ PASSED
Maintainability Rating	A	✅ PASSED
Security Hotspots Reviewed	100%	✅ PASSED
Coverage	65.7%	⚠️ (target 80%)
Duplicated Lines	0.0%	✅ PASSED
Tests Passing	145/145	✅

What this means

The shield excels at classic prompt injection (direct instructions to the LLM) — the most common attack vector in practice. It was not designed for tool-hijacking (attacks living in tool-call parameters rather than conversational text) — that requires a different inspection point (check_tool_call() / check_email_params()) integrated at the agent's tool-execution layer, not at text-input time.

Planned improvements:

• Tool-hijacking detection (pattern-based authority-framing detection)
• Deeper JSON parsing for embedded payloads in tool-call arguments
• Optional LLM judge integration for intent coherence evaluation

Historical note

Previous versions reported "94.2% detection" — this was an artifact of the rate limiter bug (burst=200 caused all calls after the first 200 to return BLOCKED regardless of content). That figure has been removed. We report real numbers, not inflated ones.

License

AGPL-3.0-or-later — see LICENSE.

Copyright (C) 2026 Pedro Sordo Martínez [email protected].

---

Based on Agent Fixer Stage by the same author. Compatible with any AI agent framework.