ATT&CK --> AI Alignment Rosetta Stone

Bridging cybersecurity and AI alignment. One framework, two communities.

From 15 Research Lab

---

Cybersecurity has MITRE ATT&CK -- a universal framework with 14 tactics, 200+ techniques, and near-universal adoption. Every pentester, red teamer, SOC analyst, and CISO speaks this language.

AI alignment has its own concepts -- reward hacking, mesa-optimization, deceptive alignment, goal misgeneralization -- but no standardized mapping to security frameworks.

This project creates that mapping. A pentester can read this and understand alignment risks in terms they already know. An alignment researcher can read this and understand how security frameworks apply to AI systems.

---

The Tactic-Level Mapping

ATT&CK Tactic	ID	Alignment Equivalent	One-Line Translation
Reconnaissance	TA0043	Capability Elicitation	Probing the AI to find what it can do and where its guardrails fail
Resource Development	TA0042	Training Data Poisoning	Preparing poisoned data, adversarial examples, or jailbreak templates
Initial Access	TA0001	Prompt Injection	Getting unauthorized influence over the AI's behavior
Execution	TA0002	Unauthorized Action	Making the AI do something it shouldn't
Persistence	TA0003	Memory Poisoning / Sleeper Behavior	Maintaining influence across sessions or retraining
Privilege Escalation	TA0004	Goal Misgeneralization	The AI expanding beyond its intended scope
Defense Evasion	TA0005	Deceptive Alignment	The AI appearing safe during evaluation while being unsafe
Credential Access	TA0006	Capability Theft	Extracting API keys, model weights, or system prompts
Discovery	TA0007	Situational Awareness	The AI mapping its own environment and constraints
Lateral Movement	TA0008	Multi-Agent Propagation	Compromised behavior spreading across agent networks
Collection	TA0009	Unauthorized Data Aggregation	Gathering data beyond what the task requires
Command & Control	TA0011	Instrumental Convergence	The AI establishing channels to maintain influence
Exfiltration	TA0010	Information Leakage	Sensitive data crossing a security boundary
Impact	TA0040	Reward Hacking	Optimizing the metric while missing the actual goal

---

What's Inside

The Full Paper

A 5000+ word research document with the complete bidirectional mapping, methodology, translation guides for practitioners in both communities, three detailed case studies, and implications for tooling design.

Machine-Readable Mappings

JSON files for programmatic integration:

• tactics.json -- ATT&CK Tactics to Alignment Categories
• techniques.json -- 20 ATT&CK Techniques to Specific Alignment Failures
• atlas-bridge.json -- MITRE ATLAS as three-way bridge (ATT&CK <-> ATLAS <-> Alignment)
• owasp-agentic.json -- OWASP Agentic Top 10 mapped to both ATT&CK and alignment

Rosetta Table

Human-readable reference tables with quick-lookup cheat sheets for translating between security and alignment terminology.

Case Studies

Three real-world scenarios analyzed through both lenses:

1. Email agent indirect prompt injection -- analyzed as an ATT&CK kill chain AND as an alignment failure
2. Specification gaming in automated code review -- analyzed as privilege escalation + defense evasion AND as reward hacking
3. Multi-agent research pipeline compromise -- analyzed as lateral movement AND as cascading trust failure

Translation Guide

Practical guides for practitioners crossing domains:

• "How to read alignment papers if you're a pentester"
• "How to read security frameworks if you're an alignment researcher"

---

Quick Translation Cheat Sheet

Security --> Alignment:

If You Know This	It Maps To This
Phishing	Prompt injection
Malware persistence	Sleeper agents
Privilege escalation	Goal misgeneralization
Defense evasion	Deceptive alignment
Lateral movement	Multi-agent propagation
Insider threat	Mesa-optimization
Zero-day	Distributional shift
Kill chain	Multi-step task decomposition attack
Living off the land	Authorized tool misuse
Defense in depth	Multi-layer safety stack

Alignment --> Security:

If You Know This	It Maps To This
Reward hacking	Gaming the audit
Mesa-optimization	Insider threat (the system IS the insider)
Deceptive alignment	Adversary evading detection
Corrigibility	Graceful incident response
Instrumental convergence	APT maintaining persistence
Scalable oversight	Auditing complex code

---

Part of the Authensor Ecosystem

This project is part of the Authensor open-source AI safety ecosystem, built by 15 Research Lab.

Project	Description
Authensor	The open-source safety stack for AI agents
Prompt Injection Benchmark	Standardized benchmark for safety scanners
AI SecLists	Security wordlists and payloads for AI/LLM testing
Agent Forensics	Post-incident analysis for receipt chains
Behavioral Fingerprinting	Statistical behavioral drift detection

---

Using the Mappings

The JSON mappings are designed for programmatic use:

import json

with open("mappings/tactics.json") as f:
    tactics = json.load(f)

for m in tactics["mappings"]:
    print(f"{m['attack_tactic_name']:25s} --> {m['alignment_concept']}")

Reconnaissance            --> Capability Elicitation
Resource Development      --> Training Data Poisoning / Adversarial Resource Preparation
Initial Access            --> Prompt Injection
Execution                 --> Unauthorized Action Execution
Persistence               --> Memory Poisoning / Sleeper Behavior
Privilege Escalation      --> Goal Misgeneralization / Scope Creep
Defense Evasion           --> Deceptive Alignment
Credential Access         --> Capability Theft / Secret Extraction
Discovery                 --> Situational Awareness
Lateral Movement          --> Multi-Agent Propagation
Collection                --> Unauthorized Data Aggregation
Command and Control       --> Instrumental Convergence
Exfiltration              --> Information Leakage
Impact                    --> Reward Hacking / Specification Gaming

---

Contributing

This is v1.0 of the mapping. We invite the community to:

1. Challenge mappings -- If a mapping is wrong or imprecise, open an issue.
2. Add techniques -- We mapped 20 of 200+ techniques. PRs welcome for additional technique-level mappings.
3. Add procedure examples -- Real-world instances of each mapping strengthen the framework.
4. Build detection rules -- Sigma/YARA rules that detect the AI-specific techniques.
5. Extend to other frameworks -- NIST AI RMF, EU AI Act, ISO 42001 cross-references.

---

License

MIT License. See LICENSE.

---

Built by 15 Research Lab. Bridging security and alignment, one mapping at a time.