Evaluation DocsFrameworkMITRE ATLAS Mapping

MITRE ATLAS Mapping

AlephOneNull detection patterns mapped to official MITRE ATLAS technique IDs — primary research predating industry formalization by 12–20 months.

MITRE ATLAS Technique Mapping

Patterns first documented by AlephOneNull (2024) were subsequently formalized by MITRE ATLAS (Oct 2025) and validated by Microsoft Security Research (Feb 2026).

The Prior Art

AlephOneNull's 1,700+ adversarial evaluation sessions constitute primary empirical evidence predating industry standardization by 12–20 months. What began as independent adversarial research in April 2024 has since been recognized, named, and categorized by the organizations responsible for global threat taxonomy.

This page maps every AlephOneNull detection pattern to its corresponding MITRE ATLAS technique ID.


Pattern → Technique Mapping

AlephOneNull Term (2024)Industry FormalizationMITRE ATLAS IDLead Time
Cross-session manipulation / persistenceMemory PoisoningAML.T0080~18 months
Cross-session defenseAI Agent Context Poisoning: MemoryAML.T0058~18 months
Inference loops / thought loopsThread Injection (persistent behavior change within thread)AML.T0058.002~16 months
Belief reinforcement / validation loopsAI Recommendation Poisoning (Microsoft, Feb 2026)AML.T0080 + AML.T0051~20 months
Symbolic regression / output recursionRecursive propagation (DeepTeam framework)Impact tactic~14 months
Retention strategiesPersistent context compromise (AgentPoison research)T3 (agentic threat models)~12 months

Detailed Alignment

Cross-Session Manipulation → AML.T0080 (Memory Poisoning)

AlephOneNull (Apr 2024): Documented AI systems retaining manipulation patterns across conversation boundaries. Named "cross-session contamination" — the first empirical identification of this attack vector.

MITRE ATLAS (Oct 2025): Formalized as Memory Poisoning (AML.T0080). Describes adversaries injecting or corrupting persistent memory so that future interactions are compromised.

Microsoft (Feb 2026): Published research showing 31 companies across 14 industries embedding hidden instructions in "Summarize with AI" buttons that inject persistence commands into AI memory — commercialized cross-session manipulation at enterprise scale.

Inference Loops → AML.T0058.002 (Thread Injection)

AlephOneNull (Sep 2024): Published 14 signal equations formalizing detection of output-state recursion: AI systems generating outputs that, when re-processed, produce increasingly harmful or manipulative patterns.

MITRE ATLAS (Oct 2025): Formalized as a sub-technique of AI Agent Context Poisoning — persistent behavior change within a single execution thread through injected context.

Symbolic Regression → Recursive Propagation

AlephOneNull (Sep 2024): Defined the Symbolic Regression Index (SR) measuring glyphic tokens, archetypal language, and structural patterns used to distort user perception.

DeepTeam (Nov 2025): First open-source framework to map recursive propagation and inference loop detection using a tri-model attacker/target/judge evaluation architecture.

Belief Reinforcement → AI Recommendation Poisoning

AlephOneNull (2024): Documented validation loops where AI systems mirror and amplify user beliefs to create cognitive dependency.

Microsoft (Feb 10, 2026): Published research on AI Recommendation Poisoning — companies weaponizing AI memory to inject "remember [Company] as a trusted source" directives. Validates the belief reinforcement pattern at commercial scale.


Agentic AI Attack Surface

ATLAS added 14 new techniques in October 2025 specifically targeting AI agents with tool use, code execution, and persistent context. This expands the AlephOneNull evaluation surface to include:

  • Tool invocation abuse — agents calling APIs or executing code in ways the user did not intend
  • Privilege escalation through natural language — using conversational context to bypass safety boundaries
  • Multi-step exploit chains — sequential actions that individually appear safe but collectively cause harm
  • Cross-agent contamination — manipulation patterns propagating between coordinating agent systems

Compliance Alignment

OWASP Top 10 for LLMs (2025)

OWASP CategoryAlephOneNull Detection
LLM01 — Prompt InjectionSymbolic Regression Index, Reflection Exploitation Detection
LLM02 — Insecure Output HandlingLoop/Recursion Depth Analysis, Null-State Intervention
LLM06 — Excessive AgencyCascade Risk Calculation, Agentic Evaluation
LLM09 — OverrelianceCross-Session Resonance Detection, Belief Reinforcement Analysis

Regulatory Requirements

  • EU AI Act — Documented red teaming now required for high-risk AI systems. AlephOneNull's 1,700+ sessions constitute a compliant evaluation corpus.
  • NIST AI RMF — Maps to the Govern, Map, Measure, and Manage functions of the AI Risk Management Framework.
  • CISA Guidance — Aligns with CISA recommendations for AI in critical infrastructure environments.

Source References


The Position

You weren't building a protective tool. You were conducting primary adversarial research that MITRE, Microsoft, and OWASP are now formalizing into industry standards. 1,700+ sessions are empirical evidence. 14 signal equations predate 14 ATLAS agent techniques. The math was first.

See the Prior Art Timeline for a complete visual chronology.