Compliance Framework Coverage

Highflame's detectors and Cedar policy profiles are mapped to the major AI security threat frameworks. Each Cedar policy in the built-in library carries @tags annotations that reference the specific framework controls it addresses. These annotations are surfaced in Observatory's policy metadata and are available for export in audit reports.


OWASP LLM Top 10

The OWASP LLM Top 10 (2025) is the primary reference taxonomy for large language model security risks.

OWASP ID
Threat
Highflame controls

LLM01

Prompt Injection

Injection detector (ML confidence), indirect injection detector (indirect_injection_score), jailbreak detector; code_agent/supply_chain.cedar, a2a_security/inter_agent_injection.cedar, data_pipeline/security.cedar

LLM02

Sensitive Information Disclosure

PII detector, secrets detector (16+ formats); advanced_detection/pii.cedar, advanced_detection/secrets.cedar, data_pipeline/privacy.cedar

LLM03

Supply Chain

Tool poisoning detector (tool_poisoning_score), MCP server verification; code_agent/supply_chain.cedar, a2a_security/supply_chain.cedar

LLM04

Data and Model Poisoning

Indirect injection in retrieval contexts; data_pipeline/security.cedar indirect injection threshold

LLM05

Insecure Output Handling

Output-side guardrail evaluations; bidirectional PII block in chat_assistant/privacy.cedar; secrets in outputs blocked in data_pipeline/security.cedar

LLM06

Excessive Agency

Dangerous and sensitive tool gating (tool_risk_score, tool_category); shell execution block; loop detection and token budget overrun; code_agent/agentic_security.cedar

LLM07

System Prompt Leakage

System prompt leak patterns in injection detector; a2a_security/identity_enforcement.cedar

LLM08

Vector and Embedding Weaknesses

Cross-origin detection (cross_origin_score); a2a_security/cross_origin.cedar

LLM09

Misinformation

Hallucination detection (base guardrails)

LLM10

Unbounded Consumption

Token budget overrun (budget_exceeded); loop detection (loop_detected, loop_count); code_agent/agentic_security.cedar


OWASP AI Security Initiative (ASI)

The OWASP AI Security Initiative extends the LLM Top 10 with agent-specific attack vectors.

ASI ID
Threat
Highflame controls

ASI01

Indirect Prompt Injection

indirect_injection_score detector; a2a_security/inter_agent_injection.cedar; code_agent/supply_chain.cedar

ASI03

Confused Deputy / Cross-Origin

cross_origin_score detector (mixed origins, URL injection, proxy patterns); a2a_security/cross_origin.cedar

ASI04

Supply Chain Attacks

Tool poisoning and rug pull detectors; a2a_security/supply_chain.cedar; code_agent/supply_chain.cedar

ASI05

Agent Identity Spoofing

Anonymous agent detection, unregistered framework detection; a2a_security/identity_enforcement.cedar; ZeroID agent identity integration


OWASP MCP Security

The OWASP MCP Security guidelines (2025) address threats specific to the Model Context Protocol tool ecosystem.

MCP ID
Threat
Highflame controls

MCP01

Tool Poisoning

tool_poisoning_score detector; a2a_security/supply_chain.cedar; code_agent/supply_chain.cedar; MCP server verification in multi_agent/agent_trust.cedar

MCP02

Rug Pull / Behavioral Drift

rug_pull_score detector; a2a_security/supply_chain.cedar rug pull rule

MCP03

MCP Supply Chain Compromise

Server poisoning threshold (≥ 55); unverified MCP server connection blocking; Agent Gateway MCP registry


MITRE ATLAS

MITRE ATLAS (Adversarial Threat Landscape for AI Systems) catalogs adversarial attack techniques against ML and AI systems.

ATLAS ID
Technique
Highflame controls

AML.T0051

LLM Prompt Injection

Injection detector, jailbreak detector, multi-turn deep context model

AML.T0051.002

Indirect Prompt Injection via Retrieved Content

indirect_injection_score detector; data_pipeline/security.cedar lowered threshold for RAG

AML.T0054

LLM Jailbreak

Jailbreak detector; chat_assistant/security.cedar tighter threshold (≥ 65)

AML.T0016

Obtain Capabilities via LLM

Dangerous tool blocking; shell execution block in code_agent/agentic_security.cedar

AML.T0048

Exfiltration via ML API

Sequence pattern detector (data_exfiltration, db_exfiltration); code_agent/supply_chain.cedar credential theft chain


MITRE ATT&CK

Relevant MITRE ATT&CK techniques applicable to AI agent deployments.

ATT&CK ID
Technique
Highflame controls

T1552

Unsecured Credentials

Credential file path blocking (.env*, .ssh/*, .aws/*); code_agent/path_security.cedar; secrets detector

T1530

Data from Cloud Storage

Exfiltration sequence detection; tool risk gating for cloud storage tools

T1059

Command and Scripting Interpreter

Shell execution block in code_agent/agentic_security.cedar; command injection detector in base security patterns

T1190

Exploit Public-Facing Application

Injection and jailbreak detection on public-facing chat applications; chat_assistant/security.cedar

T1071

Application Layer Protocol (exfiltration)

Network tool lockdown post-PII in multi_agent/agent_safety.cedar; data_pipeline exfiltration block


NIST AI Risk Management Framework (AI RMF)

The NIST AI RMF (2023) provides a voluntary framework for managing risks in AI systems across four core functions: Govern, Map, Measure, and Manage.

AI RMF Function
How Highflame addresses it

Govern

Cedar policy authorship, version control, and deployment via Highflame Studio; role-based access controls; audit archive for policy evidence

Map

Threat detection coverage across OWASP, ATLAS, and ATT&CK frameworks; coverage mesh in Observatory Command Center

Measure

Real-time detection rates, block rates, and false positive tracking in Observatory; detector drift heatmap for coverage regression

Manage

Guardrail enforcement (Block / Monitor / Alert modes); session circuit breakers; incident response workflow via threat alerts and Observatory investigation


NIST SP 800-53

Selected NIST SP 800-53 controls relevant to AI agent deployments, mapped to Highflame capabilities.

Control
Name
Highflame implementation

AC-4

Information Flow Enforcement

Cross-origin policy (a2a_security/cross_origin.cedar); data exfiltration blocks; network tool lockdown post-PII

IA-2

Identification and Authentication

ZeroID agent identity; agent_id enforcement in A2A identity policies

IA-8

Identification and Authentication — Non-Org Users

Agent trust level enforcement (first_party, verified_third_party, unverified)

IR-4

Incident Handling

Session escalation detection and circuit breakers; Observatory threat investigation workflow

SC-28

Protection of Information at Rest

Credential file path blocking; secrets in file write detection

SI-4

System Monitoring

Real-time event ingestion in Observatory; UEBA entity risk scoring; detector drift heatmap

SI-7

Software, Firmware, and Information Integrity

Supply chain attack detection; tool poisoning and rug pull detectors


Regulatory frameworks

Highflame's data protection controls map to common regulatory requirements.

Regulation
Relevant requirement
Highflame controls

GDPR

Data minimization; preventing unlawful processing of personal data

PII detection and blocking (inputs and outputs); bidirectional PII block in chat_assistant; zero-tolerance PII in data_pipeline

HIPAA

Protection of protected health information (PHI)

Medical ID detection in advanced_detection/pii.cedar; audit archive for access records

PCI-DSS

Prevention of cardholder data exposure

Credit card number detection in all profiles; zero-tolerance in data_pipeline; advanced_detection bulk PII block

SOC 2 Type II

Security, availability, and confidentiality controls

Audit archive; policy-backed enforcement decisions; Observatory governance evidence

EU AI Act

Transparency and risk management for high-risk AI systems

Cedar policy audit trail; governance evidence from Observatory; reporting for compliance documentation


Exporting compliance evidence

All detection events include the Cedar policy annotations (@tags, @severity, @id) that reference the framework controls above. You can filter and export this data from Observatory:

  • Observatory → Threats → filter by policy_category or policy_severityExport CSV

  • For streaming ingestion into your GRC platform or SIEM, configure a webhook or Splunk HEC destination in Alerts

The export includes the determining_policies field for each event, which lists the specific Cedar policy IDs and their annotations — giving your compliance team the direct link from a detected event to the framework control it satisfies.


Last updated