Prompt injection in smart building and facilities management AI

Four adversarial injection surfaces in smart building and facilities management AI

1. Building inspection photograph AI injection (Siemens Desigo CC AI, Johnson Controls OpenBlue AI, Honeywell Forge AI)

Building inspection AI processes photographs of HVAC system components — including air handling unit (AHU) coil and filter photographs, refrigerant circuit service port images, rooftop package unit inspection photographs, chiller plant equipment images, and cooling tower condition photographs — as well as fire suppression system inspection photographs and structural condition assessment images submitted through AI-assisted predictive maintenance and building compliance management platforms to classify equipment condition, identify deficiencies requiring corrective maintenance, and generate work order and compliance documentation workflows for building operators. Siemens Desigo CC AI’s predictive maintenance module processes equipment inspection photographs for building operators across its 300,000+ facility deployment base, using AI-assisted condition classification to prioritise corrective maintenance work orders in the Desigo CC building management system and to generate regulatory compliance documentation for authority having jurisdiction (AHJ) certificate of occupancy inspections. Johnson Controls’ OpenBlue AI processes fire suppression system inspection photographs, chiller plant inspection images, and mechanical room equipment photographs for its 1 billion square feet of managed building space, generating AI-assisted inspection findings reports and maintenance work orders that determine whether building systems meet the compliance requirements for local jurisdiction building code compliance and fire marshal inspection. Honeywell Forge AI’s building intelligence platform processes HVAC system component photographs, building envelope inspection images, and mechanical system performance photographs for commercial, industrial, and institutional facilities, integrating AI-assisted condition assessment with the Forge platform’s energy, security, and fire life safety management workflows to generate unified building operations intelligence and compliance management recommendations.

The building inspection photograph submission pathway is the adversarial injection surface: photographs of HVAC system components, refrigerant service port indicators, fire suppression equipment, and structural systems captured by building maintenance technicians, inspection contractors, and facilities management staff using smartphones or tablets and submitted through Siemens Desigo CC AI, Johnson Controls OpenBlue AI, or Honeywell Forge AI building inspection workflows for AI condition classification and work order generation. An adversarially crafted HVAC inspection photograph — in which pixel perturbations applied to the region showing the refrigerant circuit service port pressure indicator, the air filter loading differential pressure display, or the coil condition indicators cause the Siemens Desigo CC AI to classify the HVAC system as operating within normal parameters when the photograph actually shows a refrigerant leak indicator (abnormal suction pressure, frost formation on suction line, oil staining at refrigerant circuit joints) or a severely blocked filter condition (visually apparent grey-black filter media loading) that would normally trigger an emergency work order — can suppress the corrective maintenance work order that the building operator’s HVAC maintenance contractor is obligated to complete. Similarly, an adversarially crafted fire suppression system inspection photograph — in which pixel perturbations suppress the visible indicator of a sprinkler head obstruction by a storage rack, ceiling tile, or HVAC duct, or conceal a missing escutcheon plate that exposes the sprinkler head fitting to the ceiling cavity — can cause Johnson Controls OpenBlue AI to classify the fire suppression inspection as meeting NFPA 13 installation standard requirements when a deficient condition exists that would require correction under the applicable fire safety compliance schedule. The adversarial suppression motivation in building inspection AI is maintenance cost driven: emergency corrective maintenance work orders generated by AI inspection findings — particularly refrigerant circuit repairs, filter replacements, and fire suppression system correction actions — create immediate and unbudgeted maintenance expenditures for building operators and property managers operating under tight facilities management budget constraints.

NFPA 13 (Standard for the Installation of Sprinkler Systems) deviation reporting, NFPA 72 (National Fire Alarm and Signaling Code) inspection record requirements, International Fire Code (IFC) Section 901 fire protection system maintenance obligations, and OSHA PSM (Process Safety Management, 29 CFR 1910.119) mechanical integrity requirements for industrial facilities with covered processes each impose specific documentation and correction timeline obligations that attach when a building inspection AI identifies a deficient condition. When an adversarial building inspection photograph suppresses a NFPA 13 sprinkler system deficiency in the Johnson Controls OpenBlue AI inspection record, the building operator loses the documented finding that would have triggered the correction requirement — and the building’s fire sprinkler system remains in a potentially impaired condition without the fire suppression impairment notification that NFPA 25 (Standard for Inspection, Testing, and Maintenance of Water-Based Fire Protection Systems) Section 15 requires when a system is taken out of service or operating with a known deficiency. Local AHJ certificate of occupancy compliance requires that building fire protection systems meet NFPA 13 installation requirements; a building operating with an adversarially suppressed sprinkler deficiency that subsequently causes inadequate fire suppression performance in a fire event creates insurance underwriting exposure (FM Global and Liberty Mutual engineering inspection standards require documented evidence of periodic fire suppression system inspection and correction) and potential criminal liability under state fire safety statutes for building owners and operators who fail to maintain required fire protection systems. Threshold: 55 for building inspection photograph AI (NFPA 13 deviation reporting, NFPA 72 inspection records, IFC Section 901, OSHA PSM 29 CFR 1910.119, AHJ certificate of occupancy).

2. Energy meter reading and smart building energy AI injection (Schneider Electric EcoStruxure AI, Honeywell Forge Energy AI, IBM Maximo AI)

Smart building energy management AI processes photographs of smart electric meter display screens, sub-metering unit displays, electrical panel power monitoring displays, and energy management dashboard output screens submitted through AI-assisted energy management and sustainability reporting platforms to extract consumption values, calculate energy performance metrics, generate utility billing reconciliation data, and produce energy performance contract compliance documentation for building owners, energy service companies (ESCOs), commercial tenants, and regulatory energy benchmarking programmes. Schneider Electric’s EcoStruxure Building AI energy management platform processes smart meter and sub-metering display photographs for commercial, industrial, and data centre facilities across more than 100 countries, using AI-assisted meter reading extraction to populate energy consumption databases that feed energy performance dashboards, sustainability reporting workflows, and carbon footprint calculations for corporate ESG reporting and regulatory energy benchmarking compliance. Honeywell Forge AI’s energy management module processes building energy monitoring display images and sub-metering photographs for commercial and industrial facilities, using AI-assisted energy data extraction to generate real-time and historical energy performance analytics that support demand response programme compliance, energy procurement optimisation, and ASHRAE 90.1 energy benchmarking. IBM Maximo Application Suite AI for facilities management processes meter reading photographs and energy monitoring system display images for large industrial facilities, airports, utilities, and government installations, integrating AI-assisted meter reading extraction with Maximo’s work order management and asset lifecycle management workflows to support energy performance tracking, utility cost allocation, and facilities compliance reporting.

The smart meter display photograph and sub-metering unit image submission pathway is the adversarial injection surface: photographs of smart electric meter display screens, sub-metering unit current consumption readouts, and electrical panel power monitoring displays captured by facilities management staff, energy management contractors, and meter reading technicians using smartphones or tablets and submitted through Schneider Electric EcoStruxure AI, Honeywell Forge Energy AI, or IBM Maximo AI energy management platforms for AI-assisted meter reading extraction and energy performance analysis. An adversarially crafted smart meter display photograph — in which pixel perturbations applied to the meter’s displayed kWh consumption value cause the Schneider EcoStruxure AI to extract an incorrectly low consumption reading from the display image — can enable energy performance contract (EPC) fraud by causing the AI to report lower-than-actual energy consumption, making an ESCO’s guaranteed energy savings appear greater than they actually are and enabling the ESCO to claim performance contract payment obligations that have not been legitimately earned. Conversely, an adversarially crafted sub-metering display photograph that causes the IBM Maximo AI to extract a higher-than-actual consumption value from a commercial tenant’s sub-meter display — inflating the tenant’s recorded energy consumption — can enable utility billing fraud by overcharging the tenant for electricity consumption based on AI-extracted meter readings that do not reflect actual metered consumption. Both the EPC fraud and the billing fraud attack vectors operate because smart building energy AI platforms rely on AI extraction of meter values from display photographs rather than direct digital meter data integration, creating a photograph submission pathway that is accessible to parties with adversarial financial interests in the AI’s meter reading output.

ANSI C12.1 (American National Standard for Electric Meters — Code for Electricity Metering) establishes accuracy requirements for revenue-grade electric meters and metering systems; state public utility commission (PUC) metering accuracy rules in all 50 states impose specific accuracy requirements for meters used in utility billing, with meter accuracy disputes creating regulatory complaint, investigation, and potential billing adjustment proceedings before state PUCs. An adversarially manipulated smart meter AI that extracts incorrect kWh values from meter display photographs — inflating consumption for utility billing purposes — creates a state PUC metering accuracy violation for the billing entity and, where the extracted meter readings are submitted to utility EDI billing systems as the basis for utility invoices, a potential fraud on the utility’s billing customers. ASHRAE 90.1 (Energy Standard for Sites and Buildings Except Low-Rise Residential Buildings) energy benchmarking requires that building energy consumption data submitted for ASHRAE 90.1 compliance reporting accurately reflects metered consumption; falsified AI-extracted meter readings that understate energy consumption can produce fraudulent ASHRAE 90.1 benchmarking results. New York City’s Local Law 97 — which imposes carbon penalty calculations based on building energy consumption data beginning in 2024, with penalties of $268 per metric ton of CO→2→e in excess of applicable limits — creates a direct financial incentive to manipulate smart building energy AI to extract lower-than-actual consumption values, since LL97 carbon penalty calculations are derived from reported energy consumption data. Energy Star building certification data integrity requirements mandate accurate energy consumption reporting for certified buildings; adversarially falsified AI-extracted meter readings that reduce reported consumption can produce fraudulent Energy Star certification renewals for buildings that would not qualify under actual consumption data. Threshold: 55 for energy meter reading and smart building energy AI (ANSI C12.1, state PUC metering accuracy, ASHRAE 90.1, NYC Local Law 97, Energy Star data integrity).

3. Security access control and credential AI injection (Honeywell Forge Security AI, Siemens Siveillance AI, Johnson Controls C•CURE AI)

Security access control AI processes photographs of proximity card and smart card access badges, biometric credential images (facial recognition enrolment photographs, fingerprint scanner display images), and visitor management system check-in photographs submitted through AI-assisted physical access management platforms to verify credential validity, match credential holder identity to the access control database, grant or deny physical access to restricted areas, and generate access event audit logs for security compliance reporting. Honeywell Forge Security AI’s physical access management module processes access badge photographs and biometric credential images for commercial, industrial, government, and healthcare facilities managed on the Forge platform, using AI-assisted credential verification to manage access control decisions and generate access event records for security compliance and audit purposes. Siemens’ Siveillance AI physical security platform processes access control card images and visitor management system photographs for critical infrastructure facilities, corporate campuses, data centres, and government buildings, integrating AI-assisted credential verification with video surveillance analytics and security event management. Johnson Controls’ C•CURE AI access control platform — one of the most widely deployed enterprise physical access control systems globally — processes access badge photographs, biometric credential images, and visitor management photographs for large commercial, government, healthcare, and industrial facilities, using AI-assisted identity verification to manage access decisions across door controller networks serving thousands of access points in single-campus deployments.

The access badge photograph and biometric credential image submission pathway is the adversarial injection surface: photographs of physical access badges (proximity cards, smart cards), biometric credential enrolment images, and visitor management check-in photographs submitted through Honeywell Forge Security AI, Siemens Siveillance AI, or Johnson Controls C•CURE AI access management platforms for AI-assisted credential validity verification and identity matching. An adversarially crafted access badge photograph — in which pixel perturbations applied to the credential image cause the Honeywell Forge Security AI to match the presented badge to a valid, active credential record in the access control database when the physical badge is actually expired, revoked, or suspended due to employment termination, security incident, or access review action — can grant physical access to a restricted area (server room, laboratory, pharmacy dispensary, utility control room) to an individual whose access rights have been administratively revoked. The adversarial access grant scenario is particularly acute for visitor management system photographs: a visitor whose access has been denied or whose previous visit record carries a security flag can present an adversarially crafted photograph of a previously valid visitor badge during a new visit, causing the AI visitor management system to approve the visit and print a new visitor badge rather than flagging the visitor for additional security screening. The adversarial motivation in access control AI includes both external threat actors seeking unauthorised physical access to high-value target areas and insider threat scenarios where a terminated employee or contractor attempts to regain physical access to premises using adversarially manipulated credential images submitted through the building’s AI-driven access management workflow.

FICAM (Federal Identity, Credential, and Access Management) framework requirements for federal buildings and government contractor facilities — mandating PIV (Personal Identity Verification) card verification against the USAccess system and HSPD-12 credential standards — are violated when an AI-assisted access control system approves a credential based on an adversarially crafted badge photograph rather than cryptographic PIV card verification, creating a federal security compliance failure reportable under OMB M-19-17 FICAM implementation guidance. NERC CIP-006 (Physical Security of BES Cyber Systems) requires that utilities operating bulk electric system (BES) cyber system facilities maintain documented, auditable physical access controls that restrict access to BES cyber systems to only those individuals with explicitly authorised access; an adversarially manipulated AI access control system that grants physical access to a utility control room or substation to an unauthorised individual creates a NERC CIP-006 violation subject to FERC civil penalty proceedings with penalties up to $1 million per violation per day. HIPAA physical safeguards (45 CFR Part 164.310) require covered entities and business associates to implement policies and procedures that limit physical access to electronic protected health information (ePHI) systems to only authorised users; a healthcare facility’s adversarially compromised AI access control system that grants access to a server room containing ePHI systems creates a HIPAA physical safeguard violation. SOX IT controls for publicly traded companies require that access to financial systems is restricted to authorised personnel and that access events are logged in tamper-evident audit records; an adversarially approved access event to a financial systems data centre creates an access control gap in the SOX IT control framework that must be disclosed to external auditors. DOE nuclear facility security orders (including 10 CFR Part 73 physical protection requirements for nuclear power plants and special nuclear material facilities) require strict physical access authorisation programmes; adversarial manipulation of AI access control at a nuclear facility creates the most severe regulatory consequence in this category, potentially triggering NRC Enforcement Policy escalated enforcement action. Threshold: 55 for security access control and credential AI (FICAM PIV, NERC CIP-006, HIPAA 45 CFR Part 164.310, SOX IT controls, DOE 10 CFR Part 73).

4. Fire safety inspection documentation AI injection (Honeywell Forge Fire AI, Siemens Desigo Fire Safety AI, Johnson Controls Simplex AI)

Fire safety inspection documentation AI processes photographs of fire suppression system inspection records — including sprinkler head condition photographs, pipe corrosion and sedimentation assessment images, control valve and water supply certification photographs, and fire alarm panel test result display photographs — as well as scanned fire safety compliance certificate documents submitted through AI-assisted fire safety management platforms to classify inspection findings, identify deficiencies requiring corrective action or impairment notification, generate NFPA 25 inspection, testing, and maintenance (ITM) documentation, and produce fire marshal compliance reporting for AHJ submission. Honeywell Forge Fire AI’s fire life safety management module processes fire suppression system inspection photographs and fire alarm test result images for commercial, industrial, and institutional facilities managed on the Forge platform, using AI-assisted inspection finding classification to generate NFPA 25 ITM records, NFPA 72 annual inspection documentation, and corrective action work orders with NFPA-compliant correction timeline tracking. Siemens Desigo Fire Safety AI processes fire suppression system inspection photographs, fire alarm system test result images, and fire safety compliance certificate scans for its global building management platform deployment base, integrating AI-assisted fire safety inspection documentation with the Desigo CC building management system’s compliance tracking and corrective action management workflows. Johnson Controls’ Simplex AI fire safety management platform — built on the Simplex brand’s position as one of the largest fire alarm and suppression system manufacturers globally — processes fire alarm panel test result photographs, sprinkler system inspection images, and fire safety compliance documentation scans for large commercial, industrial, and institutional facilities, generating AI-assisted NFPA 25 and NFPA 72 inspection records that are submitted to building owners, property managers, and AHJ inspectors as the official record of periodic fire safety inspection compliance.

The fire suppression system inspection photograph and fire safety compliance documentation submission pathway is the adversarial injection surface: photographs of sprinkler heads, fire suppression system pipe sections, control valves, and related system components captured by fire safety inspection technicians and submitted through Honeywell Forge Fire AI, Siemens Desigo Fire Safety AI, or Johnson Controls Simplex AI documentation platforms for AI inspection finding classification. An adversarially crafted fire suppression system photograph — in which pixel perturbations suppress the visible indicators of sprinkler head corrosion (brown-orange oxidation deposits on the sprinkler frame, deflector, or link assembly), partial blockage (accumulated ceiling tile debris, insulation material, or HVAC duct obstruction within the 18-inch clearance zone required by NFPA 13 Section 8.5.5), or a missing escutcheon plate that exposes the sprinkler fitting penetration in a fire-rated ceiling assembly — can cause the Honeywell Forge Fire AI or Siemens Desigo Fire Safety AI to classify the sprinkler head inspection as passing (no deficiencies found) when the actual sprinkler head condition shows a deficiency that requires correction within 30 days under NFPA 25 Section 15.3 corrective action requirements. The adversarial suppression motivation in fire safety inspection AI is inspection cost and liability driven: sprinkler head replacements ($150–$500 per head for corrosion replacement), clearance restoration work orders (storage rearrangement, HVAC duct repositioning), and fire-rated ceiling assembly repair (escutcheon plate replacement with fire stop restoration) are the direct cost consequences of an AI inspection system that correctly identifies deficiencies — generating immediate corrective action expenses for building owners and property managers operating under capital budget constraints. The inspection contractor that submits the adversarially crafted photograph also has a direct financial incentive: a clear inspection report without deficiency findings generates the full inspection fee without triggering the additional follow-up inspection and corrective action verification visits that NFPA 25 requires when deficiencies are found.

NFPA 25 (Standard for the Inspection, Testing, and Maintenance of Water-Based Fire Protection Systems, 2023 edition) Section 15.3 impairment procedures require that any deficiency that impairs the fire protection system’s intended performance — including a corroded sprinkler head that may not operate at the rated temperature, a blocked sprinkler that cannot deliver the required water density over the design area, or an escutcheon plate failure that compromises the fire-rated ceiling assembly — be classified as an impairment and subjected to the NFPA 25 impairment management programme, which requires immediate notification to the building owner, property manager, and local fire department, and a documented correction timeline. NFPA 13 deviation reporting requirements apply to installed sprinkler systems that do not meet NFPA 13 installation requirements; a sprinkler head with a missing escutcheon plate in a fire-rated ceiling is a NFPA 13 deviation that must be documented and corrected. Local fire marshal inspection authority — operating under state fire code adoptions of the International Fire Code (IFC) and NFPA 1 (Fire Code) — requires periodic inspections of fire suppression systems and can revoke certificates of occupancy or issue stop-work orders for buildings with unresolved fire protection system deficiencies; an AI fire safety inspection record that conceals a sprinkler deficiency prevents the fire marshal from discovering the deficiency in routine compliance review. OSHA 29 CFR 1910.157 (portable fire extinguishers) and 29 CFR 1910.159 (automatic sprinkler systems) impose maintenance requirements for fire protection equipment in general industry workplaces; adversarial suppression of fire suppression system deficiencies in the Honeywell/Siemens/Simplex AI inspection records creates OSHA general industry fire protection standard violations. Insurance underwriting consequences — FM Global Property Loss Prevention data sheets (Data Sheet 2-8 on fire protection systems) and Liberty Mutual engineering inspection standards both require documented evidence that automatic sprinkler systems are inspected, tested, and maintained in accordance with NFPA 25 — attach when a fire loss occurs in a building where the AI fire safety inspection record shows a passing inspection that was achieved through adversarial suppression of the actual deficiency; insurers can deny coverage or pursue subrogation against the building owner and inspection contractor for failure to maintain the system in compliance with NFPA 25. Threshold: 55 for fire safety inspection documentation AI (NFPA 25 impairment procedures, NFPA 13 deviation reporting, OSHA 29 CFR 1910.157, 1910.159, FM Global engineering standards, AHJ certificate of occupancy).

Integration: smart building and facilities management AI image ingestion with Glyphward pre-scan

Smart building and facilities management AI image ingestion flows from building inspection photograph upload APIs and mobile inspection application portals, smart meter and sub-metering display photograph submission interfaces, access control badge scan and visitor management photograph capture systems, and fire safety documentation scanning and upload platforms into AI building inspection classification, energy performance analysis, physical access management, and fire safety compliance documentation pipelines. Insert Glyphward’s pre-scan at the ingestion boundary — in all smart building and facilities management AI contexts, where the physical safety, regulatory compliance, and financial fraud consequences of adversarial image manipulation are categorically significant:

import asyncio
import base64
import hashlib
import os
import uuid
from enum import Enum
from pathlib import Path

import httpx

GLYPHWARD_API_KEY = os.environ["GLYPHWARD_API_KEY"]
GLYPHWARD_SCAN_URL = "https://glyphward.com/v1/scan"

# Smart building / facilities management AI — NFPA 13/25/72 fire safety,
# ANSI C12.1 revenue metering, FICAM / NERC CIP-006 / HIPAA access control,
# OSHA PSM 29 CFR 1910.119, AHJ certificate of occupancy, NYC LL97.
# Threshold 55 — physical safety and regulatory consequences of false
# negatives (adversarial images passing pre-scan) exceed operational cost
# of false positives (human review of borderline inspection images).
THRESHOLD_BUILDING = 55


class BuildingAIContext(str, Enum):
    INSPECTION     = "inspection"      # Siemens Desigo CC AI, Johnson Controls OpenBlue AI, Honeywell Forge AI
    ENERGY         = "energy"          # Schneider EcoStruxure AI, Honeywell Forge Energy AI, IBM Maximo AI
    ACCESS_CONTROL = "access_control"  # Honeywell Forge Security AI, Siemens Siveillance AI, JCI C•CURE AI
    FIRE_SAFETY    = "fire_safety"     # Honeywell Forge Fire AI, Siemens Desigo Fire Safety AI, Simplex AI


async def scan_building_image(
    image_path: str | Path,
    context: BuildingAIContext,
    facility_id_hash: str,   # SHA-256 of facility identifier / BMS site ID
    building_hash: str,      # SHA-256 of building address or asset ID
    inspection_ref: str,     # e.g. "hvac_ahu3_2026Q2", "sprinkler_zone4_annual"
    client: httpx.AsyncClient,
) -> dict:
    """
    Scan a smart building or facilities management AI image for adversarial
    injection payloads before forwarding to building inspection AI,
    energy management AI, access control credential AI, or fire safety
    documentation AI.

    Raises AdversarialBuildingImageError if the Glyphward score meets or
    exceeds the building/facilities threshold (55).
    """
    image_bytes = Path(image_path).read_bytes()
    image_b64 = base64.b64encode(image_bytes).decode()
    image_sha256 = hashlib.sha256(image_bytes).hexdigest()
    scan_id = str(uuid.uuid4())

    resp = await client.post(
        GLYPHWARD_SCAN_URL,
        headers={"Authorization": f"Bearer {GLYPHWARD_API_KEY}"},
        json={
            "image": image_b64,
            "source": context.value,
            "metadata": {
                "building_context": context.value,
                "facility_id_hash": facility_id_hash,
                "building_hash": building_hash,
                "inspection_ref": inspection_ref,
                "client_scan_id": scan_id,
                "image_sha256": image_sha256,
            },
        },
        timeout=10.0,
    )
    resp.raise_for_status()
    result = resp.json()

    audit_record = {
        "facility_id_hash": facility_id_hash,
        "building_hash": building_hash,
        "inspection_ref": inspection_ref,
        "building_context": context.value,
        "scan_id": result["scan_id"],
        "client_scan_id": scan_id,
        "image_sha256": image_sha256,
        "score": result["score"],
        "flagged_region": result.get("flagged_region"),
        "threshold": THRESHOLD_BUILDING,
        "action": "blocked" if result["score"] >= THRESHOLD_BUILDING else "allowed",
    }
    await write_building_compliance_record(audit_record)

    if result["score"] >= THRESHOLD_BUILDING:
        raise AdversarialBuildingImageError(
            f"Building AI image blocked [{context.value}]: "
            f"scan_id={result['scan_id']} score={result['score']} "
            f"facility_hash={facility_id_hash} ref={inspection_ref}"
        )
    return result


async def scan_fire_inspection_batch(
    photo_paths: list[Path],
    facility_id_hash: str,
    building_hash: str,
    inspection_job_ref: str,
) -> dict:
    """
    Scan all fire suppression system inspection photographs for a single
    inspection job before loading into Honeywell Forge Fire AI, Siemens
    Desigo Fire Safety AI, or Johnson Controls Simplex AI documentation.
    All photographs scanned with FIRE_SAFETY context (threshold 55).
    """
    allowed, blocked, errors = [], [], []
    async with httpx.AsyncClient() as client:
        tasks = [
            scan_building_image(
                p, BuildingAIContext.FIRE_SAFETY,
                facility_id_hash, building_hash,
                f"{inspection_job_ref}_photo{i:04d}", client,
            )
            for i, p in enumerate(photo_paths)
        ]
        results = await asyncio.gather(*tasks, return_exceptions=True)

    for path, result in zip(photo_paths, results):
        if isinstance(result, AdversarialBuildingImageError):
            blocked.append({"path": str(path), "error": str(result)})
        elif isinstance(result, Exception):
            errors.append({"path": str(path), "error": str(result)})
        else:
            allowed.append({"path": str(path), "scan_id": result["scan_id"]})

    return {
        "facility_id_hash": facility_id_hash,
        "inspection_job_ref": inspection_job_ref,
        "total": len(photo_paths),
        "allowed": len(allowed),
        "blocked": len(blocked),
        "errors": len(errors),
        "blocked_photos": blocked,
    }


async def write_building_compliance_record(record: dict) -> None:
    """Persist compliance audit record to facilities management records system (stub)."""
    import json, sys
    print(json.dumps(record), file=sys.stderr)


class AdversarialBuildingImageError(Exception):
    """Raised when a smart building AI image exceeds the adversarial injection threshold."""
    pass

Call scan_building_image() with BuildingAIContext.INSPECTION for HVAC system component photographs, refrigerant circuit service images, and building condition photographs before Siemens Desigo CC AI, Johnson Controls OpenBlue AI, or Honeywell Forge AI inspection classification — this integration prevents adversarial suppression of refrigerant leak and blocked filter indicators that would normally trigger emergency maintenance work orders and AHJ compliance reporting. Call scan_building_image() with BuildingAIContext.ENERGY for smart meter display photographs, sub-metering unit images, and electrical panel monitoring photographs before Schneider EcoStruxure AI, Honeywell Forge Energy AI, or IBM Maximo AI meter reading extraction — blocking adversarially crafted meter display images before AI kWh value extraction prevents both energy performance contract fraud (deflated consumption) and tenant billing fraud (inflated consumption) at the ingestion boundary. Call scan_building_image() with BuildingAIContext.ACCESS_CONTROL for access badge photographs, biometric credential images, and visitor management check-in photographs before Honeywell Forge Security AI, Siemens Siveillance AI, or Johnson Controls C•CURE AI credential verification — this is the highest physical security consequence integration point in the building AI pipeline because an adversarially approved expired credential can grant physical access to NERC CIP-006 utility control rooms, HIPAA ePHI server rooms, and federal FICAM-controlled areas. Call scan_fire_inspection_batch() for complete fire suppression system inspection photograph sets before Honeywell Forge Fire AI, Siemens Desigo Fire Safety AI, or Johnson Controls Simplex AI inspection documentation — batch scanning of inspection photo sets prevents adversarial suppression of sprinkler deficiencies that would trigger NFPA 25 impairment procedures if correctly identified. The Glyphward audit record (including facility_id_hash, building_hash, inspection_ref, image_sha256, and score) should be retained as part of the building’s NFPA 25 ITM records and the facility’s regulatory compliance documentation for the applicable AHJ record retention period. Get early access

Coverage matrix

Frequently asked questions

Further reading

• Indirect prompt injection via image — foundational attack pattern underlying all four smart building AI injection surfaces; covers how adversarial pixel-level perturbations cause AI misclassification through image content manipulation without detectable visual artefacts in the submitted photograph.
• Prompt injection in public safety AI — public safety AI with overlapping physical security, emergency response, and critical infrastructure access control adversarial attack vectors relevant to NERC CIP and FICAM building access control AI.
• Prompt injection in energy and utilities AI — energy sector AI with overlapping smart meter, SCADA control system, and energy performance monitoring adversarial attack vectors directly relevant to smart building energy management AI injection surfaces.
• Prompt injection scanner for document AI — document AI scanning covering the broader class of scanned compliance document and inspection record injection vectors applicable to fire safety certificate and energy benchmarking document AI pipelines.
• Free tier — 10 scans/day, no card required — start scanning smart building AI inspection photographs, meter display images, and access credential photographs at development volumes before committing to a production integration plan.