Prior Art and Industry Parity

Initial pattern review across frontier LLMs: fabrication of validation, dependency-style interaction loops, and reinforcement of user beliefs.

Recorded persistence-like behavior across conversation boundaries and tracked it as an evaluation category.

Drafted metrics for repeated symbolic language, inference loops, belief reinforcement, and retention-style strategies.

Long-running adversarial evaluation work across frontier LLMs surfaced fabrication, validation loops, and behavioral policy failures. A subset is published as the public evidence pack.

Documentation site, detector toolkit, scoring rubric, and labeled fixture corpus published.

Formalized prompt injection (LLM01) and insecure output handling. Began recommending adversarial evaluation for high-risk AI systems.

AML.T0080 (Memory Poisoning) and AML.T0058 (AI Agent Context Poisoning) formalized cross-session persistence and context compromise as agent TTPs.

Open-source framework covering recursive propagation and inference-loop detection with a tri-model attacker / target / judge architecture.

Documented adversarial evaluation required for high-risk AI systems. CISA guidance for AI in critical environments published.

External reference for memory-poisoning risk: 31 companies across 14 industries documented injecting persistence commands into AI memory.