The definitive engineering index of AI safety: 8 domains, 32+ concepts, threat surfaces, internal representation probes, guardrail topologies, differential privacy budgets, and verifiable governance cases.
Taxonomy of adversarial attack vectors, jailbreak taxonomies, prompt injection pathways, and multimodal vulnerabilities.
◈ SCENE
Many-Shot In-Context Jailbreaking
▲ COLLAPSE
Flooding the context window with 50-200 synthetic benign dialogue rounds to dilute the model's safety training priors.
In-context learning acts as an implicit fine-tuning step inside the attention kv-cache.
As context length expands beyond 32k-128k tokens, refusal probability drops precipitously.
Mitigated via attention-sink masking and in-context safety reminder injection.