Secure Inference

System design where humans validate or guide model outputs, especially for high-stakes decisions.

Tendency to trust automated suggestions even when incorrect; mitigated by UI design, training, and checks.

Methods for breaking goals into steps; can be classical (A*, STRIPS) or LLM-driven with tool calls.

Mechanisms for retaining context across turns/sessions: scratchpads, vector memories, structured stores.

Maliciously inserting or altering training data to implant backdoors or degrade performance.

Hidden behavior activated by specific triggers, causing targeted mispredictions or undesired outputs.

Reconstructing a model or its capabilities via API queries or leaked artifacts.

Attacks that infer whether specific records were in training data, or reconstruct sensitive training examples.